Content uploaded by Dulani Meedeniya
Author content
All content in this area was uploaded by Dulani Meedeniya on Mar 24, 2021
Content may be subject to copyright.
Received February 23, 2021, accepted March 10, 2021, date of publication March 15, 2021, date of current version March 23, 2021.
Digital Object Identifier 10.1109/ACCESS.2021.3065965
Glioma Survival Analysis Empowered With
Data Engineering—A Survey
NAVODINI WIJETHILAKE 1, (Student Member, IEEE),
DULANI MEEDENIYA 1, (Member, IEEE), CHARITH CHITRARANJAN1,
INDIKA PERERA 1, (Senior Member, IEEE), MOBARAKOL ISLAM2, (Member, IEEE),
AND HONGLIANG REN 3,4, (Senior Member, IEEE)
1Department of Computer Science and Engineering, University of Moratuwa, Moratuwa 10400, Sri Lanka
2Biomedical Image Analysis Group, Imperial College London, London SW7 2AZ, U.K.
3Department of Biomedical Engineering, National University of Singapore, Singapore 117583
4Department of Electronic Engineering, The Chinese University of Hong Kong, Hong Kong
Corresponding author: Dulani Meedeniya (dulanim@cse.mrt.ac.lk)
This work was supported by the University of Moratuwa, Sri Lanka under the Conference & Publishing Grant.
ABSTRACT Survival analysis is a critical task in glioma patient management due to the inter and intra tumor
heterogeneity. In clinical practice, clinicians estimate the survival with their experience, which can be biased
and optimistic. Over the past decades, diverse survival analysis approaches were proposed incorporating
distinct data such as imaging and genetic information. The remarkable advancements in imaging and high
throughput omics and sequencing technologies have enabled the acquisition of this information of glioma
patients efficiently, providing novel insights for survival estimation in the present day. Besides, in the past
years, machine learning techniques and deep learning have emerged into the field of survival analysis of
glioma patients trading off the traditional statistical analysis-based survival analysis approaches. In this
survey paper, we explore the prognostic parameters acquired, utilizing diagnostic imaging techniques and
genomic platforms for survival or risk estimation of glioma patients. Further, we review the techniques,
learning and statistical analysis algorithms, along with their benefits and limitations used for prognosis
prediction. Consequently, we highlight the challenges of the existing state-of-the-art survival prediction
studies and propose future directions in the field of research.
INDEX TERMS Survival prediction, risk analysis, glioma, genomics, radiomics, radiogenomics, prognosis.
I. INTRODUCTION
A. GENERAL OVERVIEW OF GLIOMAS AND SUBTYPES
Gliomas are tumors that occur in glial or other progenitor
cells. Gliomas account for 26.7% of all primary brain and
central nervous system tumors. Generally, gliomas occur
within the brain, mostly in the frontal, parietal, temporal
lobes, and rarely in the occipital lobe. They also develop in
spinal code and cauda equina cerebrum [1]. Based on the his-
tology, Glioblastomas (GBM) account for 54.7% of primary
brain and central nervous systems gliomas. Astrocytomas and
GBMs combined account for 75% of gliomas. The incidence
rate of gliomas decreases with the increasing age for the
children and adolescents (age between 15-19 years), and
approximately 46.5% of tumors are gliomas in this particular
age group [2]. However, the incidence rate increases for the
The associate editor coordinating the review of this manuscript and
approving it for publication was Hao Luo .
patients over 20 years and the highest is among the age
of 85+years. The median age at the diagnosis of GBM
is 65 years. Further, GBM, astrocytomas, and oligoden-
drogliomas are comparatively higher in males with light skin
colours than females [3].
The initial glioma classifications consider the underlying
histology based morphological appearances in particular cell
types. The brain cells have intensive networks that help to
maintain the functions of the human brain [4].
•Astrocytes: These are the main connective tissue cells
that can be found in the brain. When these cells show
morphological similarities in a specific region, they are
called astrocytomas.
•Oligodendrocytes: These cells wrap the neuronal axons
in the brain with a myelin sheath. Oligodendrogliomas
derive from these cells.
•Ependymal cells: Ependydoma occurs in these cells,
which occurs less frequently in humans.
43168 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ VOLUME 9, 2021
N. Wijethilake et al.: Glioma Survival Analysis Empowered With Data Engineering—A Survey
World Health Organization (WHO) has categorized the
astrocytomas, oligodendrogliomas, oligoastrocytomas, and
ependymal gliomas into four grades based on the natural
disease cause, absence, and presence of anaplastic features
and malignancy [5].
•WHO grade I: Least malignant behavior and slow-
growing. Pilocytic Astrocytomas belong to this.
•WHO grade II: Slow growing, but sometimes, a brain-
invasive growth can be seen. Diffuse astrocytoma
belongs to this category.
•WHO grade III: Rapidly growing gliomas, with histo-
logical features of anaplasia.
•WHO grade IV: Most malignant glioma, known as
GBM and distinguished by the presence of necrosis and
microvascular proliferation.
Recently, the underlying molecular pathogenesis is ana-
lyzed to identify genetic alterations, which can cause gliomas.
This is slightly complementary to histological classifications
and diagnostics.
The combination of WHO grade II and III gliomas are
often referred to as Lower Grade Gliomas (LGG). Although
the traditional classification of gliomas based on histology
separates glioma into astrocytomas, oligodendrogliomas, and
mixed oligoastrocytomas, these three can be categorized into
2 subtypes based on their molecular profiles.
•Mutation in tumor protein p53 (TP53) gene along with
the Mutation in α-thalassemia/mental retardation syn-
drome X-linked (ATRX) gene: These gene markers indi-
cate ’astrocytic’ genotype.
•Mutation in telomerase reverse transcriptase (TERT)
promoter and co-deletion of chromosomal arms 1p and
19q: These gene markers indicate the ’oligodendroglial’
genotype.
Therefore, the existence of oligoastrocytomas histological
type is not required any more [6]. Both of these genotypes fre-
quently show mutations in Isocitrate dehydrogenase 1 (IDH1)
and sometimes in Isocitrate dehydrogenase 2 (IDH2). Fur-
ther, the novel classification of gliomas is into three groups
based on the outcomes, natural histories, and the response for
the treatments [4].
•IDH1 mutant, 1p/19q co-deleted tumors are mostly
oligodendroglial histology: these tumors have the best
prognosis/ longer survival.
•IDH1 mutant, 1p/19q non co-deleted tumors that are
mostly astrocytic morphology: these tumors have an
intermediate survival
•IDH wild-type, 1p/19q non co-deleted tumors: these
have a poor prognosis/ short survival.
Gliomas, based on their histology, localization, and
growth, have different levels of mortality and morbidity.
GBM, the most malignant glioma, has only a median survival
of 15 months, even with surgical resection followed by radi-
ation therapy and chemotherapy [7]. In fact, less than 3% of
GBM patients survived after 5 years from the initial diagno-
sis [8]. Moreover, WHO grade II and WHO grade III gliomas
have median overall survival times of 78.1 and 37.6 months,
respectively [9]. Considering the histology, the median sur-
vival of astrocytomas and oligodendrogliomas are 5.2 and
7.2 years, respectively [10].
B. IMPORTANCE OF SURVIVAL ESTIMATION I N GLIOMA
Overall survival of glioma patients is measured either from
the date of diagnosis of glioma or from the date the treatments
begin. The length of the time is measured until the patient
is alive. This measurement is frequently used to decide the
impact of a particular treatment given to the patient. The clin-
ical protocols are planned to maximize the overall survival
always without compromising the quality of life of patients.
The quality of life of a glioma patient relies on several facts,
such as the location and size of the tumor, complications that
occur due to surgical interventions, effects of radiotherapy
and pharmacotherapy,and psychosocial consequences, which
ultimately impact the survival of patients [4]. These fac-
tors are associated with impaired neurocognitive functions,
communication abilities, and acute toxicities and functional
deficits. Long term surviving patients, treated with radiother-
apy, have a high risk for neurocognitive impairments [11].
Moreover, standard treatments such as anti-seizure agents and
corticosteroids, used in routine medications, cause nausea,
fatigue, drowsiness, and many more short term side effects.
Meanwhile, they can induce long term effects such as diabetes
mellitus, depression, hypertension, and osteoporosis [4].
Moreover, other than physical consequences, non-physical
impairments arise in spiritual, psychological, behavioral
domains. Nevertheless, the families get affected by the illness
and side effects causing vulnerable psychological trauma
and financial consequences. Therefore, it is essential for the
clinicians, neurosurgeons, and oncologists to decide what
treatments the patients require for long term survival with
minimum side effects. Sometimes, the patients will weaken
due to side effects than the effects of the glioma. More often,
clinicians estimate the survival of patients by clinical fac-
tors assessment, imaging assessment, and their experience.
Researches show that these decisions can be biased, inaccu-
rate, and optimistic [12]. This will also impact the patient
and their families to face the situation and prepare for future
challenges.
Further, the resources such as medications, therapies are
required to be allocated among the patients based on the
requirement giving priority to the prognosis of each patient.
Thus, accurate survival prediction is a predominant factor
for better treatment planning and clinical management with
optimal utilization of resources [13].
C. CURRENT ISSUES
Traditionally, the research on survival prediction of gliomas
with medical images and computational modelling is mostly
performed with the acquisition of images, such as Mag-
netic Resonance Images (MRI), Computed Tomography
(CT), or Positron Emission Tomography (PET), followed
by preprocessing, feature extraction, and classification into
short, medium and long survival groups [14], [15]. Initially,
VOLUME 9, 2021 43169
N. Wijethilake et al.: Glioma Survival Analysis Empowered With Data Engineering—A Survey
semi-automated approaches were popular, where segmen-
tation and shape, histogram, and volume feature extrac-
tion were performed manually, and machine learning (ML)
based classifications were performed for survival estimation.
State-of-the-art, fully automated deep learning (DL) based
segmentation followed by feature extraction and ML based
classification and regression models are the most com-
mon approaches with a maximum accuracy reported
around 70% [16]. These imaging modalities are still unable
to capture the intra-tumor heterogeneity, which occurs at the
molecular level; although, imaging approaches have enlight-
ened the survival prediction research through its non-invasive
behaviour.
Mostly in gliomas, statistical analysis and risk estimation
approaches are proposed to estimate the survival and risk.
The risk is given as a risk score that involves expression
values of few genes, or any other factors related to survival
and even can be measured using nomograms. However, these
approaches are also unable to accurately reflect the survival
of glioma patients. Also, there are various methodologies
proposed by different research teams, that are independent
and unique, making it more complicated to compare their
approaches and apply in the clinical domain.
D. MOTIVATION
The new improvements in the high throughput technologies
such as microarray and DNA/RNA sequencing have made
the acquisition of heterogeneous genomic features credi-
ble and thus, more accurate survival analysis can be con-
ducted, overcoming the drawbacks in the traditional radiomic
approaches [17]. Nevertheless, the advancements in fields
such as Artificial Intelligence, with the capability of han-
dling high dimensional data and models, has arisen in the
field of patient management related to cancer [18]. This can
provide an absolute orientation for the future research of
glioma prognosis estimation. Therefore, to incorporate the
novel approaches, it is necessary to be aware of the current
state-of-the-art techniques while identifying their ambiguities
and inadequacies.
In addition, the current exploration is not yet feasible to be
implemented in glioma clinical management. Through this
work, we were motivated to explore the current approaches
for survival prediction of gliomas, focusing on radiomics,
genomics and other survival related data. Consequently,
we identify the limitations and challenges that might gravitate
the state-of-the-art survival prediction approaches for a clini-
cally plausible era. Thus, we believe this survey will motivate
many researchers to find directions to produce more precise
patient survival estimation models.
E. CONTRIBUTION
This study mainly explores the current survival analysis tech-
niques such as ML algorithms, DL approaches and statistical
analysis techniques. Our first objective is to identify the dif-
ferent types of data, including imaging features and genomic
data, used in the survival prediction of gliomas. Moreover, we
compare the preprocessing and other techniques by ana-
lyzing the accuracy and limitations in the existing studies.
As another objective, we analyze the ML and DL algorithms
that have been used for survival prediction with distinct data.
Next, we review the statistical analysis based methods, apart
from learning techniques, that are being used for survival or
risk assessment of glioma patients. Ultimately, we report the
limitations of the current research that can be addressed in
the future studies of glioma survival prediction, providing an
explicit perception of clinical execution.We suppose that this
research will broaden the horizons of the field of survival
prediction of glioma patients forging ahead to an effectual
glioma management.
This study is important for researchers who are engaged
in developing tools and algorithms for survival estimation of
glioma and various other cancer types to get a general idea
about the state-of-the-art methods. Other than that, this is
important for clinicians to decide the capability of utilizing
the available algorithms in clinical practice, and to provide the
directions required for the development in this field, further
adding clinical value and importance.
II. DATA TYPES USED IN GLIOMA SURVIVAL ANALYSIS
A. GLIOMA SCREENING AN D DATA COLLECTION IN
CLINICAL PRACTICE
The screening of gliomas is frequently done using several
types of tests. Physical examining is the initial test done
by clinicians, where the medical history, family history and
lifestyles are assessed closely. As neurological tests, a set
of questions are asked from the patients to determine the
changes that have occurred in the brain, spinal cord and
nerves with the progression of the tumor [19]. Visual field
tests are also commonly performed in the initial diagno-
sis process, since eyesight can get disrupted with the mus-
cles being compressed due to the tumor within the brain.
After these types of initial screening normally imaging tests
are obtained. The most frequently used imaging tests are
Computed Tomography Scan (CT Scans), Magnetic Reso-
nance Imaging (MRI), Single Photon Emission Tomogra-
phy Scan (SPECT), Positron Emission Tomography (PET)
and Angiograms. Other than that, blood and urine tests are
done to identify the substances, that change as an effect of
tumor [19].
However, the cancerous status of the tumor has to be
determined by examining the cells using a microscope. For
this, biopsy or surgical resection are performed. In biopsy,
only a piece of tissue from the tumor region is acquired and
in surgical resection, the complete tumor is removed with the
help of images acquired from the tumor region [20]. With the
tumor tissue sample received with one of the above methods,
the pathologists closely observe the cells, i.e. the pathological
images, through Light & Electron Microscopes, to verify the
signs of cancer and further, genomic profiles are acquired.
These genomic profiles include the gene expression, methy-
lation and mutation profiles, which we further discuss
in Section II-C.
43170 VOLUME 9, 2021
N. Wijethilake et al.: Glioma Survival Analysis Empowered With Data Engineering—A Survey
FIGURE 1. Overview of this survey paper on glioma prognosis.
Consequently, all these types of test profiles are achieved
from glioma patients, throughout the diagnosis and screen-
ing processes. Thus, many tests are collected from each
patient before diagnosis and after the surgery, and these
records are collected by many institutions for research and
various other purposes. These records are used to develop
various prognostic tools and algorithms, to provide precise
personalized disease management as we discuss in this work.
Some institutions have made these details available to the
public with open access as explained in Section II-E.
B. RADIOMICS
As discussed in Section II-A, different imaging techniques
are used for clinical purposes in practice. However, for
VOLUME 9, 2021 43171
N. Wijethilake et al.: Glioma Survival Analysis Empowered With Data Engineering—A Survey
FIGURE 2. A general radiomics extraction pipeline.
TABLE 1. Open-source Tools available for Pre-processing and Segmentation of MRI.
survival related analysis these images have to be processed
to extract imaging features from the tumor region. In this
section, we discuss the process of acquiring those features
step by step. A general pipeline for extracting radiomics
from MRI sequences is shown in Figure 2. This has 4 main
steps; 1) image acquisition: MRI scans are acquired, 2) image
preprocessing: the images are skull stripped, registered and
normalized, 3) image segmentation: the tumor region is seg-
mented into subregions, and finally, 4) Radiomic feature
extraction: various image related features are extracted. These
four steps are explained as follows.
1) MRI ACQUISITION
The conventional state-of-the-art approach for survival pre-
diction is the radiology imaging-based method. As a
non-invasive tool in cancer clinical protocol, structural MRI
is widely used to capture morphological, functional, and
structural information related to cancer. Mostly, T1-weighted
(T1W) and T2-weighted (T2W) that shows the basic
pulse sequences in MRI, Fluid-attenuated inversion recovery
(FLAIR), and gadolinium-enhanced sequences of MRI are
used for diagnosis, surveillance, and monitoring of gliomas.
Specifically, the pathological enhancement, which is com-
mon is malignant gliomas, can be denoted by acquiring
and assessing FLAIR and contrast-enhanced T1 weighted
imaging sequences [21]. There are several publicly avail-
able datasets with MRI sequences and corresponding clin-
ical information for survival analysis, which we discuss in
Section II-E. Other than that, many institutions use their own
private MRI cohorts, collected and processed by themselves
for prognosis studies.
2) MRI PREPROCESSING
A major obstacle for performing segmentation on MRI
images is the presence of non-brain tissues, eyeballs, and
skin. Therefore, these parts should be stripped to obtain a
clear image that can be used for further exploration. There
are several skulls stripping methods, such as morphology-
based, intensity-based, deformable intensity-based, hybrid-
based, and atlas-based [22]. Further, there are software tools
such as 3D slicer [23] for skull stripping manually other
than the filter algorithms implemented in ITK [24]. Skull
stripping follows the image registration, warping, and inten-
sity normalization, as image preprocessing steps, in order
to similarize the metadata of the sequences. Lao et al. [25]
have used the open-source software ITK for skull stripping,
rigid registration and intensity normalization via histogram
matching. Nonetheless, Wijethilake et al. [26], [27] have used
the open-source software 3D slicer for skull stripping and
registration. Moreover, Tixier et al. [28] have implemented
a C++ wrapper function with Insight ToolKit for the rigid
registration of MRI for their work. In Table 1, we have
summarized the commonly used tools available for structural
MRI analysis of the brain.
3) SEGMENTATION
After image pre-processing, the aforementioned MRI
sequences are used to delineate the regions of interest of
the brain tumor into the pathological sub-regions known as
necrosis, enhancement, and edema [34]. This can be done
using both manual annotations and also automated tech-
niques. ITK-SNAP [29] is the widely used software tool
for manual segmentation. However, manual segmentation
43172 VOLUME 9, 2021
N. Wijethilake et al.: Glioma Survival Analysis Empowered With Data Engineering—A Survey
TABLE 2. Segmentation methods and imaging features extracted in glioma related studies.
is a tedious, laborious and time-consuming process. The
annotation can also vary with the observer. Therefore,
with the advancements in the field of artificial intelli-
gence, DL based automated segmentation extensively utilizes
well-known convolutional neural network (CNN) architec-
tures such as U-Net [35], [36] and fully convolutional net-
works (FCN) [37]. The U-Net [35] architecture is capable
of extracting both semantic and spatial information through
the encoder-decoder sections while capturing fine details
through the skip connections. Recently, 2D U-Net and as
well as 3D U-Net, with minor modifications, have also
shown promising improvements in glioma segmentation [38].
The attention block is a similar modifying state-of-the-art
DL module integrated to U-Net and other encoder-decoder
architectures for brain tumor segmentation [39]. In addition,
skip connections are also integrated within the attention
module, to avoid inconsistencies occur due to the fusion
of spatial-wise and channel-wise features. Further, U-Net
model, generative adversarial networks with different opti-
mizers together with probabilistic programming languages
have successfully used to segment medical images [36]. More
recently, the DL researchers argue that the attention should be
directed towards other factors that are important in optimizing
DL networks, such as the per-sample and population loss
functions and the optimizer, besides the DL architectures in
order to obtain a finer brain tumor segmentation [40].
Myronenko [41] has proposed a model with high perfor-
mances, which ranked 1st in the Brain Tumor Segmentation
(BraTS) 2018 challenge, for brain tumor segmentation task.
That study exploits a novel encoder-decoder based 3D archi-
tecture, with a large encoder and a small decoder, to extract
image features and build the segmentation, respectively. They
have added an extra decoder branch for the image reconstruc-
tion, where it is used to regularize the typical encoder, provid-
ing the ability to learn with limited datasets. Medical Image
Computing and Computer Assisted Interventions (MICCAI)
BraTS 2019 challenge 1st ranked winner, Jiang et al. [42]
have proposed a two-stage cascaded U-Net, where the second
stage consists of two decoders to boost the performance.
However, these automated segmentations do not accurately
separate the brain tumor and the normal brain tissue due to the
ambiguous boundaries and intensity variations. Hence, there
exists room for enhancements in segmentation with DL.
4) FEATURE EXTRACTION
The MRI segmentation is followed by feature extraction. The
texture features, intensity features, shape-based features are
extracted from each subregion [43]. Mean, variance, skew-
ness related to pixel intensities can also be extracted for a
particular tumor sub-region. Further, shape features, such as
area, centroid coordinates, length of axis, fractal dimensions,
that depict the complexity of the region can be procured.
Texture based features such as entropy, energy, correlation,
dissimilarity are extracted from each subregion [15]. The fea-
tures extracted from radiographic medical images are known
as radiomics. Recently, deep features are also obtained from
the MRI images; high-level features that depicts features
unachievable with traditional feature extraction tools.
In Table 2, we have summarized related studies and the
segmentation approaches with the extracted radiomic fea-
tures for each work. Sun et al. [44] have extracted features
based on grey level intensity, such as mean, median, standard
deviation, variance, maximum and minimum intensity values,
energy, and entropy values for edema, non-enhancing region,
necrosis, and whole tumor regions. Moreover, for shape fea-
tures, they have included the major and minor axis length,
surface area to volume ratio, surface area, volume, and other
features, that interprets the shape of each sub-region. The
features that characterize the texture of each subregion, such
as gray-level co-occurrence matrix features, gray-level run
length matrix features, are extracted as well. Pyradiomics
toolbox [45] has been used for feature extraction.
Sanghani et al. [14] have obtained 2D shape features, such
as bounding ellipsoid volume ratio and orientation, spherical
disproportion, and sphericity along with other 2D shape fea-
tures and texture features. Alternatively, Feng et al. [46] have
extracted volumetric features for each sub-region and inte-
grated with resection status and age for survival prediction.
Han et al. [47] have extracted handcrafted radiomic fea-
tures manually from the Region of Interest (RoI) and inte-
grated with deep features obtained from Visual Geometry
Group 19 weight layers(VGG-19) [48] model. Lao et al. [25]
VOLUME 9, 2021 43173
N. Wijethilake et al.: Glioma Survival Analysis Empowered With Data Engineering—A Survey
have proposed a similar approach for survival prediction
using CNN [49] to extract deep features. In that study, they
also have extracted handcrafted features, that include geom-
etry, texture, and intensity features obtained using Matlab in
house feature extracting program.
C. GENOMICS
Gliomas involve genetic and epigenetic changes that can
cause activation in the oncogenes and the inactivation in the
tumor suppressor genes. The technological growth in the field
of high throughput sequencing technologies has opened many
opportunities in survival analysis. Thus, prognosis analysing
studies related to glioma has manipulated different types
of molecular profiles, including gene expression profiles,
methylation profiles, and mutation profiles, that reflect those
genetic and epigenetic alterations. In the meantime, the cost
of acquiring those profiles has also decreased. As a result,
a large number of glioma patients’ genomic profiles are pro-
vided publicly as in The Cancer Genome Atlas (TCGA)1and
Chinese Glioma Genome Atlas (CGGA).2
1) GENE EXPRESSION PROFILING
For genomic expression level analysis, microarray technol-
ogy was used initially for obtaining the DNA sequences,
with Agilent, Affymetrix, and Illumina platforms. The gene
expression levels of the fragments of DNA are determined
by comparing them with a standard target sequence. Yet, this
technology is unable to identify pre-unknown genes. Solving
this drawback, next-generation sequencing (NGS) came into
the field of sequencing. Illumina HiSeq is a renowned mod-
ern platform in NGS, used for DNA and RNA sequencing.
NGS can provide a more comprehensive view of the tran-
scriptome by detecting pre-undetermined transcripts and
noncoding regions, and thus, appropriate for gene expres-
sion profile extraction [57]. Further, these high throughput
sequencing technologies such as NGS can accelerate modern
genomic and proteomic research with massive volumes of
data they generate [58].
2) METHYLATION PROFILING
Methylation is an epigenetic variation that adds a methyl
group to the CpG sites. This can regulate the transcription
of genes by acting on the particular genes promoter region,
differentiating the expression of the corresponding gene. Illu-
mina platforms, known as Infinium Human Methylation 450k
(IHM-450k) and Infinium Human Methylation 27k, are fre-
quently availed BeadChips kits for acquiring high throughput
DNA methylation profiles [59].
3) MUTATION PROFILING
In addition, mutations in cells can cause tumorigenesis.
Mutations can occur in tumor suppressor genes that repair
DNA alteration and cell division, and hence, causing
1http://cancergenome.nih.gov/
2http://www.cgga.org.cn/
uncontrollable cell division. IDH1/IDH2 mutations are iden-
tified as a prominent mutation that occurs in glioma patients,
which increase survival [60]. Mutation profiling identifies the
nonsense, missense, and silent mutations from the obtained
RNA sequences.
In survival prediction with genomics, gene expression
profiles are frequently deployed for risk estimation
of glioma patients [61], [62]. Nevertheless, associa-
tions between promoter methylation and gene expres-
sion are used for prognostic gene identification for risk
estimations [63].
4) RADIOGENOMICS
Radiogenomics is an emerging field for survival analysis of
many cancer related studies. In radiogenomics, the associa-
tions between radiomics and genomics are assessed.
Identifying imaging biomarkers that reflect the genomic
behavior is very much convenient for the clinicians and other
research personnel, as obtaining genomic profiles is an inva-
sive and tedious task and also not been often used in clinical
practice. Hence, GBM subtypes that have been divided based
on molecular level alterations are recognized and predicted
using the imaging biomarkers [27].
For survival prediction, radiogenomics are used in a couple
of studies related to glioma. Wijethilake et al. [26] radiomics
and gene expression features are fused together for sur-
vival prediction into short, medium, and long survival class
prediction leading for higher accuracy. Tixier et al. [28]
predict the survival by combining texture, shape features
extracted from segmented tumor regions as radiomics, and the
O-6-methylguanine-DNA methyltransferase (MGMT) methy-
lation, IDH1 mutation status as genomics.
D. OTHER DATA TYPES USED IN SURVIVAL ANALYSIS
1) HISTOPATHOLOGY
The pathological images are microscopic images obtained
from the tissue specimens of the glioma acquired with
biopsy or surgery. Histopathologists engage in analysing
microscopic images for diagnosis and prognosis purposes
in clinical practice. As we mentioned in the introduc-
tion, the gliomas are divided into astrocytomas, ependy-
momas, and oligodendrogliomas based on this underlying
histopathology. This histopathology behind gliomas is also
used for survival estimation of gliomas in a couple of studies.
Mobadersany et al. [64], Rathore et al. [65] have proposed a
DL based approach for survival prediction of glioma patients
with pathological images and integrates histology with gene
biomarkers to predict the outcome.
2) CLINICAL INFORMATION
Most of the studies that use the above discussed data types
have considered age as a clinical feature. The MICCAI
BraTS challenge 2018 best performing work has also
proved that age alone can predict survival with a high
accuracy [52].
43174 VOLUME 9, 2021
N. Wijethilake et al.: Glioma Survival Analysis Empowered With Data Engineering—A Survey
TABLE 3. Frequently used Glioma cohorts for survival analysis of Glioma patients.
E. PUBLIC GLIOMA COHORTS
There are several publicly available datasets with imaging,
genomic and histopathological records of glioma patients
with their corresponding clinical information. The National
Cancer Institute’s (NCI’s) Genomic Data Commons (GDC)3
is a platform that provides access to Genetic, pathological
and clinical data of The Cancer Genome Atlas (TCGA)
datasets, including both GBM and LGG data collections. The
Cancer Imaging Archive (TCIA) stores the radiological and
pathological data of the correlative TCGA cases. Other than
that, TCIA also provides access to new ongoing data collect-
ing projects such as the National Cancer Institute’s Clinical
Proteomic Tumor Analysis Consortium (CPTAC), that also
collect proteomics profiles of glioma patients. In addition,
the Multimodal BraTS 4provide pre-operative multimodal
MRI scans of GBM and LGG patients along with clinical
information. Unlike other glioma imaging cohorts, BraTS
dataset also provides manual annotations of the glioma, thus,
enabling it to be used as a benchmark dataset for glioma
segmentation tasks.
The Chinese Glioma Genome Atlas (CGGA) is a database
that consists of genomic data of Chinese Glioma patients.
This includes whole-exome sequencing, mRNA sequencing,
mRNA microarray, microRNA microarray, and DNA methy-
lation profiles for GBM and LGG cases. Further, there are
specific tools and applications that are developed to analyze
the DNA mutation landscape, mRNA/microRNA expression
profile, and DNA methylation profiles. Similar publicly avail-
able genomic data cohorts are Gene Expression Omnibus
(GEO) 5and the Ohio Brain Tumor Study (OBTS).
Summary of these frequently utilized datasets and the stud-
ies that have used them, are given in Table 3.
III. PREPROCESSING OF GLIOMA SURVIVAL
RELATED DATA
1) DATA CLEANING
Data cleaning is a vital task in survival prediction to eliminate
noise and improve the quality of data. Duplicate entries,
outliers, and missing data are some of the facts that might
3https://portal.gdc.cancer.gov
4https://www.med.upenn.edu/cbica/brats2020/data.html
5https://www.ncbi.nlm.nih.gov/geo/
reduce the quality of data. In medical information, missing
data is a frequent problem. Generally, the records with miss-
ing data are removed or assigned values by calculating the
mean or mode. The mean of the available continuous feature
values is assigned to the missing data of the corresponding
feature. If the feature values are categorical, the mode of the
values is assigned to the missing data of the particular feature.
However, according to the literature, removing the data will
reduce the statistical power, and the imputations can affect
the variance of the features [70].
2) NORMALIZATION
Normalization of features is mandatory to obtain better fit-
ting data and model. Standardization, also known as Z score
transformation, is a popular method used in forging the data
to follow a standard normal distribution with zero mean and
unit standard deviation. Min-Max normalization is another
common method that normalizes all the features to a simi-
lar range, mapping each feature’s minimum and maximum
values to a given range. Normalization is an important task in
any analysis to avoid the influence that some feature makes,
compared to the rest [18].
3) DIMENSIONALITY REDUCTION
Dimensionality reduction is essential in the medical field
to optimize the performance of the learning algorithm by
increasing the accuracy. According to Ladha and Deepa [71],
dimensionality reduction is vital to reduce the memory
requirements for storage and to increase the algorithm pro-
cessing speed. Moreover, redundant, noisy, irrelevant data
can be removed with this. Nevertheless, when utilizing the
algorithm in real-time or in the testing phase, the most nec-
essary features can be extracted, saving resources and time.
Dimensionality reduction has two techniques, known as fea-
ture selection and feature transformation.
•Feature Selection
In feature selection, the features that contribute to the
learning process are being selected by removing the
redundant features. A subset of original features is cho-
sen, increasing the efficiency and the accuracy of the
learning task. This can be done manually or using the
algorithms. Feature selection based on algorithms can
VOLUME 9, 2021 43175
N. Wijethilake et al.: Glioma Survival Analysis Empowered With Data Engineering—A Survey
TABLE 4. Feature selection and analyzing techniques used in survival prediction of gliomas with radiomics.
be discussed under 3 categories, filter methods, wrapper
methods, and embedded methods.
As a filter method, correlation-based feature selection
is used with high dimensional data as it requires less
computational power. Therefore, the correlation has
been utilized in feature selection with high dimensional
genomic data [72]. The most commonly used corre-
lation tools are Pearson correlation, Spearman’s rank
correlation, Kendall rank correlations, and intraclass
correlation. Based on the type of data, either continuous
or categorical, the correlation tool can be determined.
Lao et al. [25] obtained the best feature for the DL based
survival prediction model with the intra-class correla-
tion coefficient. Besides, Pearson’s correlation coeffi-
cient has been often used to identify the relationships
between differently methylated genes and differentially
expressed genes for prognostic risk score development
in glioma patients [66].
Wrapper feature selection methods tend to find the fea-
tures based on the predetermined learning algorithm,
requiring a high computational power compared to the
other two. Recursive Feature Elimination (RFE) [73]
is a famous wrapper method used for feature selection
in glioma survival prediction. This has been utilized
in radiomics based survival prediction along with Sup-
port Vector Machine (SVM) [14], [54]. The high time
consumption and computational power are significant
drawbacks of this wrapper method.
•Feature Transformation
In feature transformation, original features are trans-
formed into other sets of significant features. Thus,
the information in the original feature space remains in
the transformed feature space.
Principal Component Analysis (PCA) provides a clear
overview of multivariate data by increasing data inter-
pretability. This method creates new uncorrelated vari-
ables, maximizing the variance. PCA is considered as
the most widely used dimension reduction method [74].
IV. GLIOMA SURVIVAL ANALYSIS APPROACHES
Survival analysis of glioma patients is performed using var-
ious analysis methods, including ML, DL and statistical
analysis based methods. Since early 2000, the attention of
the research community has been drawn by the ML tech-
niques widely. As a result, much research was followed to
develop a precise survival prediction with learning methods
and features. Other than that, there are statistical approaches,
prognostic risk score models, and prognostic nomograms
proposed in glioma survival-related studies. In Table 4and
Table 5, we report feature extraction methods and the analysis
methods followed by several studies that have used radiomics
and radiogenomics, respectively as the input.
TABLE 5. Feature selection and analysing techniques used in survival
prediction of gliomas with radiogenomics.
A. MACHINE LEARNING BASED SURVIVAL ANALYSIS
Different ML algorithms are utilized to predict the survival
of glioma patients with a selected or filtered set of features.
Since early 2000, many ML techniques have been used to
predict the overall survival of glioma patient using radiomics,
genomics and radiogenomics. Moreover, these techniques
can be categorized into supervised, semi-supervised, and
unsupervised learning strategies, based on the availability of
the survival outcome. Supervised learning models require a
clinical outcome or the survival label of each patient case.
The supervised model learns to map between the given input
and the ground truth label via a pre-defined loss function.
1) SUPPORT VECTOR MACHINE
SVM is a frequently used supervised learning approach for
classification and regression tasks in the domain of can-
cer research [75], [76]. This has been used in overall sur-
vival prediction as a classification task, into short, medium,
and long survival groups and also as a regression task to
predict overall survival in days. SVM is appropriate for
high-dimensional data analysis, due to its ability to overcome
43176 VOLUME 9, 2021
N. Wijethilake et al.: Glioma Survival Analysis Empowered With Data Engineering—A Survey
the large dimensionality. For instance, SVMs have success-
fully applied to analyse gene expressions [76]. On the other
hand, in SVM, linear models can be extended to non-linear
models with kernels by transforming input instance space into
a high dimensional feature space. The SVM classification
function is as follows.
ˆ
8x0=sign w.x0+b(1)
where, x0, the test point, w, the coordinates of the separating
hyperplane and b, bias to the origin. SVM classifies the sam-
ples into classes through hyperplanes in a multidimensional
space.
SVM has been frequently facilitated in 2 class (short
<400 days and long > 400 days) and 3 class (short <300 days,
medium 300-450 days, long > 450 days) survival prediction
of Glioma patients with radiomics [14], [56]. Nie et al. [56]
have obtained deep features adopting a 3D convolutional
neural network and combine with clinical features for sur-
vival class, long and short survival prediction with SVM.
They observe that functional MRI (fMRI) and diffusion ten-
sor imaging (DTI) provide valuable information for survival
prediction than T1 MRI. Sanghani et al. [14] have used
a set of novel shape features that have a high contribu-
tion to the survival class prediction after selecting based
on RFE. They have obtained a high accuracy proving that
RFE, along with SVM, is an effective approach for survival
prediction.
Emblem et al. [77] have presented histograms of whole
tumor relative cerebral blood volume (rCBV) from MRI.
The authors have utilized SVM for survival prediction after
6 months, 1, 2, and 3 years after the diagnosis. They have
developed 4 separate models that returns the most prob-
able outcomes after 4 time-periods. This study has tested
the proposed models on independent data and has recog-
nized that their model is insensitive to the treatment changes
and image acquisition routine changes. They further demon-
strated that the proposed SVM based model can predict
survival with higher accuracy compared to an expert in the
clinical field. SVM has also been used for GBM subtype
classification [27], [78].
In Support Vector Regression (SVR), the algorithm learns
the best fitting linear function in the feature space, optimizing
the epsilon insensitive loss function with a regularizer. How-
ever, this is unable to work with the censored data related to
survival, as it is unable to identify the event occurrence for
censored instances. Thus, Khan and Zubek [79] has proposed
a SVR algorithm for censored data by modifying the epsilon
insensitive loss function asymmetrically.
2) SURVIVAL TREES
As a non-parametric tool, decision trees are widely used
in survival analysis due to their flexibility and the ability
to handle various data structures. Specifically, it does not
require to specify the links between the covariates and the
outcome beforehand as it can automatically developing inter-
actions. Basically, in a decision tree, covariate space is split
recursively until the nodes of the tree align with the outcome
of the covariates. These binary splits are performed with a
single covariate. If this covariate Xis continuous, a split is
formed with X≤c, where c is a constant. If this covariate X
is categorical, split is based on the criterion X{c1,c2, .., ck},
where c1,c2, .., ckare possible values of X. Initially, at the
root node, all the covariates are available, and recursive
binary splits are performed until the stopping criterion is
met. Consequently, this gives an overfitting large subtree.
Thus, the most appropriate subtree is chosen with a selection
and pruning method. At the terminal node, for classification,
the most prevalent class label is chosen, and for regression,
the training samples are averaged.
This tree structure is initially applied for survival analysis
by Marubini et al. [80] and Ciampi et al. [81]. The difference
between the traditional decision tree and the survival tree is
the choice of splitting criterion. The aforementioned decision
tree does not consider either the interactions between the
features or the censoring status. Originally survival trees
were proposed with the criterion of minimizing the homo-
geneity of each node that also minimizes the loss function.
Gordon and Olshen [82] have initiated this with the Wasser-
stein metric between the survival functions, paving the
ground for the criterion, based on the dissimilarity between
the nodes. Ciampi et al. [83] have proposed log-rank statis-
tics, likelihood ratio statistic, and Wilcoxon–Gehan statis-
tic to measure the heterogeneity between the nodes. This
method [84] has been utilized for establishing 6 prognostic
groups of malignant glioma patients [85]. Moreover, this
recursive partitioning analysis approach is being used for the
overall survival analysis of lower-grade glioma patients based
on pre-treatment factors [86]. With all these improvements,
recently censored observations are restored with expected
survival times, and the median survival tree is developed with
L1 loss function [87].
The recent survival studies of glioma patients exploit sur-
vival trees with various types of data related to prognosis.
Gandia et al. [88] have assessed the metabolomic profiles
acquired from biopsies to predict survival class (short, inter-
mediate, and long OS) of glioma patients with a tree-based
model. This classification tree has split into 3 branches,
where the first node criterion is the myo-inositol level and
suggests that high myo-inositol levels can cause longer sur-
vival. The second branch cleaves the patients with a low
myo-inositol level, based on the glycerol-phosphorylcholine
(GPC) level. The final branch splits based on the ala-
nine and glycine levels of glioma patients. However, this
study is limited to 46 patient cases; this might have caused
overfitting.
Fig 3represents a similar tree developed for survival class
prediction based on the gene expression levels. The first
branch of the tree is divided based on the expression level
of the Solute Carrier Family 30 Member 7 (SLC30A7) gene.
The patients with expression level less than -0.067 is split
again based on the expression of the YTH domain-containing
family protein 2 (YTHDF2) gene.
VOLUME 9, 2021 43177
N. Wijethilake et al.: Glioma Survival Analysis Empowered With Data Engineering—A Survey
FIGURE 3. Sample survival tree for survival group classification into
short, medium, and long survival.
3) RANDOM SURVIVAL FOREST
Traditional survival trees are considered to be unstable,
as it can give different survival functions even for small
permutations in the training set. Hence ensemble models,
bagging [89] and random forest regression [90] are proposed
to overcome the instability by reducing the variance of simple
trees.
In the Bagging algorithm, multiple versions of the boot-
strap sample [91] are obtained from the given data. Then, for
each sample, unpruned trees are built, and the final predic-
tion is acquired by taking the average over all the versions
of survival trees. Later this bagging algorithm inspired the
subsequent ensemble algorithm, random forests.
Unlike bagging, the splitting criterion of random forests
only uses a random subset of all the attributes at the internal
nodes, which gives the best prediction. Varieties of splitting
criterion have been suggested over the year, such as gen-
erating random cutting point for each selected feature and
selecting the best among them [92]. This idea is extended by
incorporating multiple cutting points for each selected feature
and comparing them [93].
Recently, random forest regression has been widely uti-
lized in radiomic based overall survival prediction [44], [52],
as a regressor for overall survival prediction of glioma
patients in days. Sun et al. [44] have achieved an accuracy
of 61% with the random forest and have considered a lim-
ited number of samples might have caused the inadequate
efficacy.
Choi et al. [94] have exploited non-imaging features such
as IDH status, age, WHO grade, and resection extent for
predicting overall survival. Subsequently, they incorporate
imaging features extracted from MRI for the survival anal-
ysis. Thus, they could observe a significant improvement in
survival prediction by adding radiomics. Puybareau et al. [52]
have derived imaging features and build 10 decision trees,
each with 3 randomly selected features. As a conventional
random forest, the final prediction is decided based on
the majority voting of the 10 decision trees. This method
MICCAI BraTS challenge 2018.
4) BOOSTING ALGORITHMS
Lately, boosting algorithms have emerged as a promis-
ing ensemble method for accurate prediction. Unlike for-
est methods boosting has a sequential approach where the
model learns to improve the performance of the predecessor.
The boosting approach has been applied for neural net-
works, and later, Freund and Schapire [95] have introduced
the AdaBoost (Adaptive Boost) algorithm solving previous
issues. At first, all the weights are set equally, and each round,
the model in series, is weighted, giving priority to the incor-
rectly classified examples. In gradient boosting machines,
in each round, weights are set based on the maximum corre-
lation of the predictors with the negative gradient of the loss
function [96]. Later, this gradient boosting algorithm is also
modified for censored data in survival analysis [97].
5) LINEAR REGRESSION
Linear regression is a frequently used method for survival
data analysis [26]. This is capable of modeling a linear
equation with the explanatory variables to output the depen-
dent variable. In survival analysis studies, features that have
associations with survival are used as explanatory variables,
and the dependent variable is mostly the survival in days
from the initial diagnosis or the percentage of survival after
a certain time-period. However, dealing with censored data
without knowing the actual event times is a challenge for
linear regression modeling [18].
6) BAYESIAN METHODS
Bayesian methods are recognized as a prevailing tool in
ML used in both classification and regression tasks. They out-
perform the other techniques by dealing with uncertainty with
prior knowledge to make subtle inference and predictions.
Bayesian approaches are also flexible by providing flexible
algorithms, characterising the latent structure and uncertainty.
Bayes theorem is the basic fundamental theorem behind the
Bayesian methods. It determines the relationship between the
posterior probability and the prior probability, that is affected
by a particular event that occurred in between. The Bayesian
posterior probability for a given feature set F, and the model
parameter , is given by,
p(|F)=p0()p(F|)
p(F)(2)
where p0(·) is the prior probability before the occurrence of
dataset F. after training, the model predicts for a given feature
set x[98],
p(x|F)=Zp(x,|F)d
=Zp(x|,F)p(|F)d(3)
There are two Bayesian methods, that are being widely
used in the glioma survival prediction [99]–[101].
•Naïve Bayes (NB) [102] Based on the Bayesian rule
(equation 2), for a given feature set (xthe probability
of belonging into class φ, the posterior probability of a
NB classifier is given by, P(φi|x). Thus, for a binary
classifier (with φ1andφ2, two posterior probabilities
(P(φ1|x)and P(φ2|x)) are calculated and if
P(φ1|x)>P(φ2|x)then xbelongs to class φ1and
if P(φ1|x)<P(φ2|x)then xbelongs to class φ2.
43178 VOLUME 9, 2021
N. Wijethilake et al.: Glioma Survival Analysis Empowered With Data Engineering—A Survey
The class is chosen arbitrarily if the posterior proba-
bilities are equal. However, this Naive Bayes method
assume a mutual independence between the feature
set, which is mostly not accurate for survival related
data. Also, the complexity is another drawback when
it comes to large datasets [103], [104]. Nonetheless, as a
white-box this is easily understandable for clinicians and
other personnel.
•Bayesian networks (BN) [105] In Bayesian networks,
features are considered as dependent on each other and
can be easily interpreted or visualized the relationships
between the features. Bayesian Neural Network (BNN)
was proposed by Wijethilake et al. [17] for over-
all survival class prediction with mRNA gene expres-
sion for glioma patients. The results highlight that the
BNN gives a higher accuracy than the other traditional
ML techniques such as SVM, RFC and ST.
Zhou et al. [100], [106] have leveraged the Naive Bayes
algorithm to predict the survival classes (long and short)
as a classification task. Nevertheless, they also use imag-
ing biomarkers extracted from tumor subregions in GBMs
to predict survival time as a regression task. The authors
have used distance features that represent the heterogene-
ity between regions for survival prediction. This study has
shown that Naive Bayes outperforms the SVM and K Nearest
Neighbor classifiers, with a selected subset of features. How-
ever, the deficient dataset of 16 cases is a constraint in this
study.
Further Naive Bayes has been utilized by Piccolo and
Frey [107] for predicting survival status after 2 years from
the diagnosis with molecular-level data.
7) ARTIFICIAL NEURAL NETWORKS
Artificial Neural Networks (ANN) are ML algorithms that
resemble the biological neural systems, which were estab-
lished by Rosenblatt in 1958 [108]. Thus, based on the neu-
ronal activities, the concept of ANN was initially established
by McCulloch and Pitts [109]. A basic ANN model, as shown
in Fig 4, consists of 3 layers; an input layer that receives
the input data, a hidden layer that processes the input data,
and an output layer that delivers the results. The hidden layer
analyzes the inputs and finds out patterns that give an output
based on the associations between the inputs and the outputs.
This is done in multiple iterations until the best output is
achieved, i.e., the error between the output and the expected
outputs has reached a minimum value. Each layer consists
of artificial nodes, with weights determined through learning
and predetermined activation function. The network retains
the learned parameters to predict output for an unseen input.
The black box behavior of the ANN is a significant drawback,
and also, training an ANN is a time consuming task. However,
with the recent developments in the computing technologies,
ANN is frequently used in survival prediction, while extend-
ing as deep neural networks. In fact, ANN does not require
a linear relationship between input and output to model the
behavior.
FIGURE 4. Artificial Neural Network consists of a single hidden layer.
In the review of the literature, Faraggi and Simon [110]
developed an ANN for survival analysis of cancer for the
first time in 1995. Recently, neural network is incorporated
with the Cox proportional-hazards model by converting the
output layer to a cox regression model, for survival analysis
with high throughput omics data in LGG and other types
of cancers [111]. Neural Network is utilized in glioma sur-
vival prediction to predict the survival time in days directly.
Islam et al. [15] have proposed an artificial neural network
to predict overall survival in glioma patients with radiomics.
In this work, the input layer consists of the radiomic features
such as geometrical features, volume features, texture fea-
tures extracted from the tumor region, and the output layer is
a single node directly predicting the overall survival in days.
B. DEEP LEARNING BASED SURVIVAL ANALYSIS
Recently, DL has drawn the attention of medical researchers
with the availability of large amounts of data. Neverthe-
less, DL models have shown improvements, by boosting the
precision and accuracy in many applications, over classical
ML paradigms, which we discussed in Section IV-A. Other
than for segmentation and for engineering deep features,
direct utilization of DL learning models are not frequently
used for outcome prediction in glioma patients. With the
recent advancements in digitized image acquisition and data
management associated with virtual microscopy technolo-
gies, survival is also predicted with pathological images in
several studies with DL.
Yousefi et al. [112] have proposed a deep neural network,
followed by a cox survival model for outcome prediction with
high dimensional genomic data. They have used deep survival
neural network as the DL algorithm. As an extension of
this study, Mobadersany et al. [64] have integrated genomics
with the histology images for outcome prediction, where
they observe an improvement in performance with both in
contrast with histology prediction alone. The CNN is known
as survival-CNN (SCNN) is inspired by the 19-layer Visual
VOLUME 9, 2021 43179
N. Wijethilake et al.: Glioma Survival Analysis Empowered With Data Engineering—A Survey
Geometry Group (VGG) architecture that returns a risk, given
to a cox proportional hazard layer to optimize the model via
back-propagation. Further, the SCNN is trained with histol-
ogy images to predict the risk, and integrated IDH mutation
status and the 1p/19q codeletion, as genomic biomarkers,
to train a 3 variable cox regression model.
Further, Rathore et al. [65] have proposed an over-
all survival classification approach with the Residual Net-
work (ResNet) architecture for pathology images. They have
shown an accuracy of 84.32% for survival class prediction,
long and short survival prediction in glioma patients. Both
of these studies have claimed the need for automated patch
extraction methods, to overcome the burdens, such as the time
that occur with manual patch extraction.
Several related studies have used the TCGA dataset. How-
ever, a major concern associated with training a DL network is
the requirement of a large amount of labeled data. This occurs
as a result of a high number of (millions of) parameters,
weights and biases in the convolutional layers, that learn to
depict the input features to the patient outcome while min-
imizing the prediction error. As we mentioned above there
are a specific set of data augmentation techniques followed
by DL researchers in the medical field to overcome the data
deficit. Randomized rotations and flipping of the training
data, contrast and brightness transformations are a few of
those methods. Thereby the training cohort can be expanded
to train the DL network to obtain a better performance.
Moreover, in spite of the intra-tumor heterogeneity,
the sampling regions of interests (patches) from the patho-
logical images allow the DL models to discern the varia-
tions effectively. This heterogeneity is visible in pathological
images, and thus, Mobadersany et al. [64] have selected the
regions of interest from each slide with the expert knowledge.
C. STATISTICAL ANALYSIS BASED SURVIVAL ANALYSIS
There are three types of statistical analysis methods for sur-
vival analysis. i) Non-parametric ii) Semi-parametric and
iii) Parametric [18]. Kaplan Meier and Nelson Aalen are the
most popular non-parametric methods used in survival anal-
ysis, despite the difficulties in interpretation. All these tools
are frequently utilized in glioma survival studies, to identify
prognostic genes, to visualize the survival function variations
between different categories or for comparisons.
1) SURVIVAL DATA AND CENSORING
The survival analysis explores the effect of individual covari-
ates and the time until an event, such as death or a spe-
cific state, reaches. This includes the partially observed data,
known as censored data, which makes survival analysis dif-
ferent from ML. Accordingly, if an event, such as death,
occurs during the monitoring, it is considered as an uncen-
sored record. Censored records do not experience an event
within the time-period and do not know the exact event
occurrence status that might or might not have happened after
the time-period [18].
2) KAPLAN MEIER ANALYSIS
In survival analysis, there are two functions depend on the
time. i) Survival function: Probability of surviving at least
up to time t; (pr(T>t)). ii) Hazard function: Conditional
probability of dying at time t, if the patient survived until
time t. Keplan-Meier Curve [113] is used to evaluate the
survival function based on the observed survival times, with-
out considering the underlying probability distribution. The
probability of surviving t0time is the cumulative probability
of surviving each t0time-periods from the beginning of the
study.
S(t0)=p1×p2×p3×. . . ×pt0(4)
where, p1is the probability of surviving the first time-
period. The probability of surviving each itime-period and up
to itime-period is obtained by,
pi=ki−di
ki
(5)
kiis the number of alive cases at the beginning of the itime-
period, and diis the number of deaths within the itime-period.
In the Kaplan-Meier analysis, the data where the censoring
occurred just before the itime-period are excluded from the
ri. Also, the time that the censoring occurred has a probability
of survival of 1.
3) LOG-RANK TEST
Log-rank test is a statistical hypothesis test initiated to com-
pare two survival curves. The test null hypothesis is that
there is no difference between the two survival curves (two
population groups).
χ2(log rank) =(O1−E1)2
E1
+(O2−E2)2
E2
(6)
where E1and E2are the expected events in 1 and 2 groups
and O1and O2are the observed events in 1 and 2 groups.
Although, the log-rank test can be used to explain the differ-
ence between survival curves of different groups, it does not
account for any other variables that can affect the survival
curve [18].
4) COX PROPORTIONAL HAZARD ANALYSIS
Cox proportional hazard model [114] is semi-parametric sur-
vival analysis approach and a multiple regression model, that
can account for many factors at once for the analysis. Hazard
function λ(t), the probability of dying at a particular time
if survived up to that time, is the dependent variable of this
model.
λ(t)=lim
1t→0+
pr(t6T<t+1t|t6T)
1t(7)
Now, for a given instance i, the available covariate vector
is zi=z1i,...,zpi. Thus, the cox model follows the
following hazard function.
λ(t,z)=λ0(t) exp(zβ) (8)
43180 VOLUME 9, 2021
N. Wijethilake et al.: Glioma Survival Analysis Empowered With Data Engineering—A Survey
TABLE 6. Related work on prognostic risk score calculation.
where the baseline hazard function λ0(t), is a non-negative
function of time for z=0. βis the coefficient vector asso-
ciated with each covariate, βT=β1, β2, . . . , βp. Between
any two instances the hazard ratio is,
λ(t,X1)
λ(t,X2)=λ0(t) exp (X1β)
λ0(t) exp (X2β)=exp [(X1−X2)β](9)
This demonstrates that the hazard ratio is a constant and does
not depend on the baseline hazard function λ0(t). Further,
the baseline hazard function is the same for all the subjects.
Thus, the cox model is a proportional hazard model. The
hazard ratio (HR), exp βiis obtained for each feature, and
based on that hazard ratio, the effect of those features for
survival is assessed. The HR >1, i.e., the value of βi>0,
implies that the ithfeature is positively associated with the
hazard function and thus, decreases the overall survival. Cox
proportional hazard model is used in glioma survival estima-
tion in many ways. Recently, the cut off value of HR is for
feature selection.
Zuo et al. [61] have selected genes for risk signature devel-
opment, based on the HR >1, obtained from the univariate
Cox model and further, to obtain the genes with the highest
impact, the multivariate cox model is utilized. Hsu et al. [115]
have also utilized a univariate cox model for identifying
survival-related genes.
D. OTHER METHODS USED IN SURVIVAL PREDICTION
1) RISK SCORE MODEL
Risk score modelling is another common approach for risk
estimation of glioma patients. Thus, the patients are separated
into the high risk and low-risk groups based on the expression
or methylation of genes or other features associated with
the survival. In order to obtain the most prominent genes
associated with survival, univariate and multivariate cox pro-
portional hazard analysis are frequently utilized [61]–[63].
The features are filtered based on the Hazard ratio (typically
HR > 1), and the p-value (for a significant coefficient p values
should be <0.05). Thus, based on the chosen gene signature,
the risk score is calculated, using the following formula.
Risk score =Xstatusgene ∗coefgene (10)
where coeffgene is the coefficient obtained from the Cox PH
hazard analysis for the corresponding gene and statusgene is
either the expression level or methylation status of the same
gene. The median risk score of the dataset is considered as the
threshold of high and low-risk separation of patients. In order
to give a better understanding of this clustering, heatmaps,
Kaplan Meier plots are visualized.
Zuo et al. [61] have used two publicly available GBM
cohorts, CGGA and TCGA, to identify the prognostic genes.
They have initially performed univariate cox regression anal-
ysis on both datasets separately and, based on cut-off values
of P<0.01 and HR>1, filtered 49 prognostic genes com-
mon both cohorts. Thereafter, the step-wise multivariate cox
regression analysis is performed to obtain the 6 genes, D79B,
MAP2K3, IMPDH1, SLC16A3, MPZL3, and APOBR as
a 6 gene risk signature. This work has identified these 6 gene
risk score as an adverse prognostic risk factor compared to the
other clinical factors such as chemotherapy, radiation therapy,
and age. Since they have used two cohorts, they claim that
they have been able to develop a more stable and reliable
model relative to the risk score model developed with a single
cohort.
A study carried out by Xian et al. [62] has extracted
non-coding gene expression profiles for glioma patients for
the risk score development, since previous studies have
shown non-coding RNAs cause cancer progression.
Apart from gene expression profiles, methylation pro-
files are also used for developing risk score models. Other
than that, radiomics are also utilized for risk estimations.
Beig et al. [69] extract radiomic features from T1 weighted
MRI to develop risk score models, separately based on the
gender for accurate risk estimation. In Table 6, we present
the summarize of several related works that have developed
prognostic risk score models.
2) NOMOGRAMS
Similar to the risk score prognosis model, another survival
estimation tool is the prognostic nomograms. These are gen-
erated to estimate the probability of survival after a specific
time-period of the diagnosis, based on the factors such as
gene expression of a selected set of genes. Wang et al. [116]
have developed a 5 gene prognostic nomogram for GBM
patients, considering the gene expression levels of OSMR,
BICDL1, SH3BP2, MSTN, and RGS14 genes. Similarly,
Gittleman et al. [68] have proposed a nomogram for LGG
patients, considering the Grade of the glioma (either grade II
or grade III), Sex, Karnofsky performance status, molecular
subtype. Neutrophil/lymphocyte ratio, age, the extent of the
resection, and the histology (GBM or LGG) are taken into
VOLUME 9, 2021 43181
N. Wijethilake et al.: Glioma Survival Analysis Empowered With Data Engineering—A Survey
account by Yang et al. [117] to develop a comprehensive
nomogram for glioma patients.
An example of a prognostic nomogram is shown
in Figure 5. In this, the expression levels of SLC30A7,
YTHDF2, TAF12, and STK40 genes are considered to deter-
mine the survival probability after 12 months and 18 months.
The points given to each gene is determined by the expres-
sion level that lies between 0 and 1, and then the total of
those points are used to determine the survival probability.
However, as we can see, higher the total points cause shorter
survival according to the given nomogram.
FIGURE 5. A sample nomogram was developed to estimate the
probability of survival after 12 months and 18 months from the diagnosis.
The expression levels of SLC30A7, UTHDF2, TAF12, and STK40 genes are
considered in this.
A summary of related works that have proposed prognos-
tic nomograms for survival probability estimation is given
in Table 7.
TABLE 7. Related work on prognostic nomograms development.
V. DISCUSSION
Survival Prediction of glioma patients is a critical challenge
for oncologists and other clinicians as it directly influences
the patient in many aspects. Over the past decades, many
research personnel has focused on developing survival pre-
diction approaches by utilizing various types of data and
techniques.
A. IMPORTANCE AND IMPACT OF GENOMICS IN
PROGNOSIS
Many studies have been conducted to identify the impor-
tance of genomics in gliomas, and also, some genes have
shown associations with survival. The following molecular
alterations are the most prominent alterations that are found
to have prognostic significance in glioma patients.
1) IDH1 MUTATION AND ITS IMPACT ON PROGNOSIS
IDH1 catalyze the oxidative decarboxylation of isocitrate to
α-ketoglutarate as [118]. This plays a vital role in cellular
protection from oxidative stress through the production of
NADPH [119]. WHO grade I astrocytomas were not iden-
tified with mutations in IDH1 or IDH2. In addition, WHO
grade II & III gliomas, i.e. LGG, mostly carry mutations in
IDH1, including the secondary WHO grade IV - GBM, which
arise from lower-grade gliomas [120].
In spite of all the studies, Yan et al. [120] have further
determined that the GBM patients with IDH1 mutations have
a median overall survival of 31 months, where IDH1 non-
mutant patients have a median survival of 15 months, sig-
nificantly lesser than the IDH1 mutant patients. Moreover,
patients with anaplastic astrocytomas have a median sur-
vival of 65 months and 20 months for IDH1 mutations and
non-mutations, respectively [120]. In fact, IDH1 mutation
is considered as a positive prognosis marker in lower-grade
gliomas [121]. Thus, researches verify that IDH1 mutation as
the most important prognostic marker despite the underlying
histology [60].
2) TP53 MUTATION
TP53 functions as a tumor suppressor protein by reg-
ulating cell division, which prevents cell division and
growing too fast or in an uncontrolled way. If a dam-
aged DNA cannot be repaired, the resultant protein of
TP53 inhibits DNA replication. Thus, mutated TP53 is unable
to inhibit the replication of genetically unstable cancerous
cells [122]. Thus, mutant TP53 provides an independent path-
way leading to gliomagenesis [123].
Many studies have revealed the association between the
poor prognosis of lower-grade gliomas and TP53 mutation,
and further analysis demonstrates that TP53 mutation is an
adverse prognostic marker for patients with astrocytic and
oligoastrocytoma histology [124]. Nevertheless, TP53 muta-
tion and expression of mutant TP53 in GBM patients show
an inverse correlation with the prognosis, and the chemosen-
sitivity for temozolomide also reduce with the mutant
TP53 [125]. However, researches have further demonstrated
that the prognostic effects of TP53 in GBM patients depend
on the clinical factor, age [126].
3) TERT PROMOTER MUTATION
TERT encodes telomerase, which maintains the length of the
telomere region in the eukaryotic chromosome. Telomere and
telomerase together play a crucial role in both tumors sup-
pressing and tumor promotion [127]. Consequently, TERT
promoter mutation is associated with increased expression of
the TERT gene [128]. Therefore, tumor promotions occur by
maintaining the telomere length through the overexpression
of the TERT gene [129]. Thus, TERT promoter mutation
occurs frequently in gliomas [130], [131]. Mutations in TERT
43182 VOLUME 9, 2021
N. Wijethilake et al.: Glioma Survival Analysis Empowered With Data Engineering—A Survey
promoters are found in 58% of primary GBM and less fre-
quently (28%) in secondary GBM [132]. TERT mutation is
also common in patients with Oligodendrogliomas [133].
Nevertheless, TERT promoter mutation shows a low sur-
vival rate in GBM patients [128], [132]. Further, recent
studies have shown that TERT promoter mutation has a
positive correlation with age and the WHO grade [134].
Studies have revealed that TERT promoter mutation, along
with a high relative telomere length, not only affects the
survival of glioma patients but also determine the resistance to
radiotherapy [135].
4) MGMT METHYLATION
MGMT encodes a protein that repairs DNA, damaged by
alkylating agents [136]. Thus, MGMT unmethylated gliomas
are resistant to alkylating chemotherapeutic agents, caus-
ing shorter survival of patients [137]. Moreover, methyla-
tion of MGMT silence the gene causing high sensitivity
for alkylating therapeutic agents and, thus, for a better
survival [138]. MGMT promoter methylated GBM patients
treated with temozolomide have a better prognosis (median
survival 36 months) than the patients without methylated
MGMT promoter (median survival 16.8 months) [28].
5) GENE EXPRESSION PROFILE ANALYSIS FOR GLIOMA
PROGNOSIS
The aforementioned different genetic aberrations in gliomas
and other environmental effects, cell division, genetic inher-
ent disrupt the cell behavior in cells. Thus, the regular cell
behavior changes and its protein production, i.e., translations
of genes get affected. These damaged cells reproduce rapidly
and turn into cancers. Therefore, analyzing gene expression
profiles exploits associations with molecular level genetic
alterations [125], [128]. In fact, genetic profiling reveals risk
factors, including protective traits such as immune responses,
allergies.
Typically, GBM is the most heterogeneous glioma, which
makes it challenging to treat patients. Overcoming this het-
erogeneity, gene expression profiles classify GBM patients
into 4 subtypes, based on the molecular pathogenesis, illus-
trating the prognosis of each subtype [139]. Gene expression
profile analysis, along with mutation profiles, are also used
to develop targeted therapies and other personalized drugs for
glioma patients, which also increase the overall survival [7].
In overall, genomics have a clear association with the
occurrence, progression and prognosis of gliomas. Thereby,
we speculate the importance of exploring genomics for the
glioma survival prediction, to develop more stable and accu-
rate platforms.
B. IMPORTANCE AND IMPACT OF RADIOMICS IN
PROGNOSIS
Radiomics have indicated intuitions to genomic and clinical
features of Glioma patients through radiomic analysis. Thus,
radiomics are being used to predict genetic alterations such as
IDH1 status, gene expression levels, and most frequently the
survival. Unlike genomics, radiomics provide a non-invasive
and low-cost approach for feature extraction and analysis.
Therefore, radiomics are used to predict the molecular sub-
types associated with the genomics without performing any
surgical intervention. Although genomics has a clear insight
on survival or the categorization of gliomas, the ability in
radiomics to predict the clinical parameters makes it easy
for the clinical practitioners to decide treatments due to the
non-invasive behavior.
Nevertheless, the genetic information obtained from the
tumor region might vary within the tumor. This is commonly
known as intra-tumor heterogeneity, which occurs at molec-
ular and histopathological levels. This limitation in related
genomic analysis can be avoided when using radiomics for
the analysis by considering tumor regions comprehensively,
allowing spatial mapping of distinct molecular level changes.
Moreover, based on these radiomics, 3 subtypes, known as
enhancing, irregular and solid, are identified, with each sub-
type comprising similar molecular alterations.
Many studies have found associations between imag-
ing features and the prognosis of glioma patients.
Hammoud et al. [34] had identified that glioma patients with
less or no necrosis and enhancement regions tend to survive
longer compared to the patients with greater volume of those
regions.
C. ADVANTAGES OF LEARNING APPROACHES TOWARDS
SURVIVAL ANALYSIS
Learning approaches can learn from the large amounts of
data, without explicitly specifying the prior assumptions,
rules and limits by humans. However, traditional statistical
methods are often based on assumptions such as the additivity
of the parameters in proportional hazard models and linear
functions [140]. Rajula et al. [140] have further showed that
there are conflicts in the assumptions of proportional hazard
method used in survival analysis studies in gastric cancer.
Learning approaches can handle a large number of predic-
tors where a small number of observations are available. For
example, using genomics, where thousands of gene profiles
are available, and only some observations, such as long vs
short survival, can be modelled using learning approaches
straightforwardly. Traditional statistical methods have to deal
with the limitation of handling only a limited number of
choices of factors or predictors.
Nonetheless, learning algorithms can analyse various
distinct data types, such as genomics, images at once, for
predicting outcomes [26]. Also, from a clinicians perspec-
tive, statistical approaches are not straightforward and imple-
mented in clinical practice is questionable and non-viable.
Also, the interpretability of the learning methods is yet
ambiguous. However, recently, the Explainable Artificial
Intelligence has drawn the attention of the medical data sci-
ence researchers, resolving the interpretability issue in the
medical community as we discuss in Section V-E1. These
constraints associated with explainability and interpretability
of the learning approaches will be addressed in the future.
VOLUME 9, 2021 43183
N. Wijethilake et al.: Glioma Survival Analysis Empowered With Data Engineering—A Survey
Among the learning approaches, DL methods have
shown more flexibility towards survival prediction. However,
the data hunger in DL approaches makes most of the survival
related studies to use only ML methods, as they do not have
thousands of parameters to be tuned as in DL paradigms.
However, this will also be solved in the coming decade with
the availability of more patient profiles, since most of the
institutes have now focused on collecting data. Nevertheless,
there are emerging deep probabilistic learning approaches,
that can overcome limitations such as uncertainty, extensi-
bility in traditional DL methods. Integration of these novel
techniques will lead the current survival analysis to a new era.
D. LIMITATIONS OF EXISTING STUDIES
1) LIMITATIONS OF SURVIVAL ANALYSIS WITH GENOMICS
•Heterogeneity
Specific improvements in the medical field are required
to integrate survival analysis, peculiarly with genomics,
into clinical management. For instance, molecular test-
ing and sequencing procedures should be standard-
ized and cost-effective to be implemented in practice.
In addition, it is a challenging task to develop common
data standards due to the complexity of molecular-level
data and the various high throughput platforms such
as next-generation sequencing and microarray used to
acquire different datasets. The technical limitations can
cause the same gene expression profiles obtained from
different platforms to be inconsistent. The microar-
ray gene expression profiling can get affected by the
cross-hybridization of the probes and the limited detec-
tion range of individual probes [141]. Hsu et al. [115]
has identified this as a drawback for the identification
of survival related genes, that makes their observed sur-
vival related genes differ from the other related stud-
ies. Wijethilake et al. [17] have also reported this as
a limitation for testing the proposed model on a sep-
arate dataset and thus, use cross-validation to verify
the performance of the proposed method for survival
prediction.
Nonetheless, in clinical practice obtaining genomic pro-
files requires the tissue samples obtained through inva-
sive surgical resection or biopsy. However, the procured
tissue sample, small biopsy specimens may not accu-
rately represent the molecular pathogenesis of the entire
tumor, which is mostly heterogeneous [4]. Due to this
heterogeneous nature in gliomas Xian et al. [62] have
claimed that a single biomarker is not sufficient for
survival prediction and further explorations are required
in identifying prognosis related biomarkers for clinical
implications.
•Real-Time Acquisition
Further, real-time sequencing of whole-genome exons,
along with gene expression profiling, is a challenging
task with the current technology. In clinical practice,
genomics is not routinely used due to the high cost and
the time.
2) LIMITATIONS OF SURVIVAL ANALYSIS WITH RADIOMICS
•Issues related to Image acquisition and feature extraction
pipeline
The datasets obtained from different institutions might
have followed different hardware and acquisition proto-
cols. This can also vary within the institutes. Therefore
the images acquired are quite challenging to generalize.
The radiomic feature extraction pipeline consists of sev-
eral steps, discussed in Section II-B that are prone to
errors. For instance, the manual annotations of the tumor
regions can also vary with the annotator and thus, impact
the features derived from the tumor regions. The hand-
crafted feature extraction formulas capture the textural
and morphological attributes of a given region based
on predefined mathematical formulas. However, these
implementations also change with the software used and
the techniques used. Most of these features are general
features used in any kind of images, without specify-
ing on medical images. Therefore, the implementation
might require slight changes to deal with heterogeneous
medical images obtained from different patient cases.
Feng et al. [46] have mentioned a wide range of MRI
protocols as a possible reason for the significant gap
in segmentation performance between training and the
testing phases.
3) COMMON LIMITATIONS IN SURVIVAL ANALYSIS
•Lack of Data
Most of the studies have used publicly available,
open-access datasets such as The Cancer Genome Atlas
(TCGA), Chinese Glioma Genome Atlas (CGGA).
Nevertheless, these cohorts also consist of less than
500 glioma patient cases, due to the limitations in time
and cost in data acquisition. Therefore in survival pre-
diction tasks are more likely to overfit and as a solution,
some studies only include most survival related features
to avoid overfitting [46]. BraTS challenge confirms the
ML algorithms ability to handle smaller datasets over
DL approaches [16]. Besides, Mobadersany et al. [64]
have also disclosed the need for large datasets in DL
paradigms for future research. In addition, some of the
studies also consider their own local datasets for the
research work. This causes issues in replicating the same
experiments in external cohorts, and also comparisons
with other related studies become critical. The missing
data in datasets also makes this limitation more critical.
Some work they have neglected the patients’ records
with missing data or have dealt with this issue, as we
mentioned in Section III-1.
•Data Scale and Computation requirements
The data, genomics, and also images require large vol-
umes and high-performance computing resources to
store and process. However, despite the rapid growth
of data volumes, the computing resources have a slow
development, posing a significant challenge in handling
extensive data. Nevertheless, the DL tasks in survival
43184 VOLUME 9, 2021
N. Wijethilake et al.: Glioma Survival Analysis Empowered With Data Engineering—A Survey
analysis, such as automated segmentation, are compu-
tationally expensive to train despite the requirement of
a massive amount of data for an accurate prediction.
A major challenge that occurs when handling genomics
is the ability to make plausible sense with large volumes
of data. Integrating these large volumes, with differ-
ent molecular interpretations, along with clinical data,
is also challenging.
•Reproducibility for clinical application
In addition, to use survival analysis with radiomics in
clinical practice, the radiomics should be able to repro-
duce and must be independent. According to the litera-
ture, these features can be influenced by many factors,
including the imaging equipment, acquisition proto-
cols, image preprocessing steps, and image reconstruc-
tion [142]. These can affect survival prediction models
or analysis tools. Hence, technical standardization has
emerged into research studies, and clinical practice by
associations such as Quantitative Imaging Biomarker
Alliance, in order to ensure the intra and inter-machine
reproducibility of radiomics [143]. Moreover, the repro-
ducibility of deep features, extracted from imaging data,
is also a critical task and, thus Han et al. [47] have
incorporated deep features with handcrafted features to
provide a higher accuracy for survival prediction.
Nonetheless, genomics data adhere to the same lim-
itation in reproducibility of curation and interpreta-
tion of genomic related analysis. Thus, organizations
such as the American College of Medical Genetics and
Genomics (ACMG) and the Association for Molecu-
lar Pathology (AMP) have published the standards and
guidelines for interpreting and reporting cancer related
sequence variant, providing improvements for the field
of precision medicine in cancer and genomic testing in
practice [144].
In contrast, clinical parameters have shown more
promising improvements in survival prediction as they
are generalizable between different institutions unlike
other data, that are subjected to variations in acquisition.
Thus, In BraTS 2018 challenge the best performing
model has observed the strongest correlation between
age and the survival, and further the third best perform-
ing survival prediction model has only utilized age for
the survival prediction as reported by Bakas et al. [16],
Feng et al. [46].
•Imbalanced Datasets
Some of the survival-related analyses are performed
as a classification task, as in short, medium, and long
survival group-wise classification, or as an analysis
study focusing on the censoring status. For the learning
models to effectively learn all the classes, the number
of patient cases in each class should be equal. Alter-
natively, the learning models are likely to give prior-
ity to the classes with more learning instances in sur-
vival prediction. There are commonly used sampling
techniques to avoid the class imbalance in datasets by
oversampling the minority class and under-sampling the
majority class. Wijethilake et al. [17] have used minority
class oversampling for overall survival prediction with
genomics to avoid the class imbalance.
•Multidisciplinary requirements
Research Personnel with technical, clinical knowledge
and expertise are necessary to develop a survival analysis
tools, efficiently utilizing the available resources. High
technical people tend to incorporate novel technologies
and facilities for these analyses, and the final outcome
can be less relevant in clinical practice. On the con-
trary, clinical researchers are likely to consider more
on clinical perspectives, in spite of the technical errors
such as robustness, multiple comparison errors. Multi-
disciplinary research groups will avoid these obstacles
and bring up technically efficient and accurate tools or
models with considerable clinical importance.
In addition, some of the discussed studies have got the
experts supervision from a distinct disciplinary, for dif-
ferent tasks such as for manual segmentation of tumor
regions [16] and for the region selection of pathology
images [64]. This is also important for the development
of useful survival analysis, with a coherent underlying
concepts.
E. FUTURE RESEARCH DIRECTIONS
In order to compensate for the aforementioned limitations,
the technological advancements, along with clinical applica-
tions, are progressed into novel advanced techniques. These
directions can be identified as follows.
1) EXPLAINABLE ARTIFICIAL INTELLIGENCE
The black box behavior of Artificial Intelligence algorithms
was questionable recently, with the strive to recognize the
importance of most relevant and how the predictions are
made. Moreover, in spite of the high accuracy, the clinicians
are doubtful to accept a survival prediction tool blindly with-
out a proper understanding. Several approaches are being
proposed to overcome this black box behavior, explaining
the underlying algorithm and the contribution of the features.
SHapley Additive exPlainations (SHAP) [145] is an explain-
able technique used to delineate the ML models following
a game-theoretic approach. Another renowned explainable
method is Saliency mapping [146], where a heat map is
visualized expressing the impact of each individual pixel
of the prediction and thus, explaining the classification or
segmentation neural network models. With this, the clinicians
can make sure the segmentation is error-free prior to feature
extraction and prediction.
2) AVAILABILITY OF BIOINFORMATICS TOOLS AND
COMPUTATION ENHANCEMENTS
The growth in the available computational frameworks and
libraries implemented in different programming languages
such as R and python has made the prediction task more
implementable at ease. Since most of these tools are also
VOLUME 9, 2021 43185
N. Wijethilake et al.: Glioma Survival Analysis Empowered With Data Engineering—A Survey
open access, the researchers tend to apply various techniques
and compare the performance, thus, analytically identify-
ing the most suitable survival prediction technique. More-
over, the evolution in Graphical Processing Units (GPUs)
has made the DL a possible task with parallel processing
[58], [147]. In addition, there is a possible research direction
to implement the classification algorithms with an in-build
parallelism [148]. Further, DL techniques can combine
with probabilistic models, Deep probabilistic programming
(DPP), to increase the efficiency and flexibility of the
computations [36].
3) PRECISION MEDICINE
Precision medicine analyzes the characteristics of the glioma
considering the genetic alterations, imaging markers, prog-
nosis of the patient to personalize the required treatments
of each person. In this scenario, the survival estimation of
glioma plays a crucial role in deciding the ideal treatment for
each individual that will consequently help to improve the
patient outcome. However, relying on genomics is genomics
for personalized medicine is challenging due to tumor het-
erogeneity. Therefore biopsies are repeatedly done as the
tumor progresses, and the personalized therapies are deter-
mined. For this, the initial overall survival or the risk esti-
mation is necessary to manage the treatments efficiently and
effectively.
4) ADVANCES IN GENOMIC TECHNOLOGIES
Next-generation sequencing has formed a conspicuous revo-
lution in the high throughput technologies, revealing ’driver’
gene alterations that occur in gliomas. The advancements
in the high throughput technologies have reduced the cost
of analyzing omics data, including genome, transcriptome,
proteome, and metabolome. Further, these technologies have
enabled gene expression, methylation, and mutation profil-
ing of a large number of genes at once. This transcriptome
analysis has led new directions in glioma prognosis analysis,
despite clinical and other non-invasive imaging factors. How-
ever, as we discussed earlier, the intra-tumor heterogeneity
is a critical constraint in survival analysis with genomics.
Ultra-deep sequencing sheds light on this by sequencing the
different regions of the same tumor [149]. Thus, in the future,
more comprehensive studies can be performed integrat-
ing genomics, transcriptomics, and epigenomics, the omics
beyond the exome.
5) DATA-PRIVATE COLLABORATIVE LEARNING METHODS
Most of the recent studies related to glioma prognosis have
shown promising improvements in research with the avail-
ability of data collected by various institutions. Thus, each
institutions tend to collect and analyse data by themselves.
However, these DL based models have shown biases and
inability to perform well with other unseen data acquired
by other institutions. In the current exploration, some works
and publicly available cohorts merge data from multiple
institutions, despite the concerns associated with the data
ownership, privacy and technical configurations. Data-private
collaborative learning methods are used in learning
paradigms recently, facilitating many institutions to collab-
orate their data without sharing them. This has several imple-
mentations such as federated learning, where the multiple
institutions train the learning model parallelly with their data
and aggregate learned models at a central server. Federated
learning has shown promising improvements for Glioma
segmentation [150]. This collaborative learning can also be
serial, by updating the ML model and passing it for the
next. Data-private collaborative learning can be used in future
work related to survival driving the precision of prediction
algorithms to a higher level, while maintaining common
acquisition protocols, common preprocessing routines and
similar data harmonization between institutions [151].
The genomic alterations associated with survival were
investigated since the 1990s, and the utilization of genomics
for survival or risk prediction is frequently seen after 2018.
Admittedly, imaging data were widely used for survival pre-
diction since early 2000. Yet, the advancements in segmen-
tation and feature extraction algorithms are mostly witnessed
after 2010. In this exploration, we have considered the most
recent survival prediction approaches followed by various
multidisciplinary research groups, in 2016-2020 period.
Furthermore, the future of the survival prediction for
glioma patients is drifting towards the utilization of high
dimensional inputs and models, whereas standardization and
generalization is a pivotal point. Multi-institutional collab-
orations with data-private collaborative learning can drive
this field more forward. The main obstacles that arise can be
stated as the various acquisition, preprocessing protocols fol-
lowed by different institutions. Even after these obstacles are
solved, the clinical deployment of survival prediction should
happen after close monitoring, testing and regularizing of the
standards in an application.
VI. CONCLUSION
In study has focused on survival analysis approaches fol-
lowed by existing studies related to gliomas, that has a
low survival compared to other cancer types. In the mod-
ern era of personalized medicine, the survival prediction of
glioma patients are performed utilizing diverse imaging and
gene biomarkers. Although this area is being improved and
drawn the attention of the clinicians in the past couple of
decades, still, there are requirements for significant improve-
ments to provide survival prediction routinely in clinical
practice. Through this survey, we have explored the potential
of genomics in glioma survival analysis, although it has an
invasive approach. We further reported the importance of
using imaging for survival prediction, due to its non-invasive
behavior and as imaging is frequently used as an initial diag-
nostic tool. In this survey, we disclosed the limitations in both
approaches and enlighten the researchers who are interested
in survival prediction of glioma patients to follow novel
strategies to create maximum insight out of miscellaneous
data. We further emphasized the importance of clinicians’
43186 VOLUME 9, 2021
N. Wijethilake et al.: Glioma Survival Analysis Empowered With Data Engineering—A Survey
feedback to develop clinically feasible survival prediction
software application. Thereby the future glioma management
will thrive with accurate survival prediction approaches, ulti-
mately provoking a better prognosis in glioma patients.
ACKNOWLEDGMENT
The authors acknowledge the support received from the Con-
ference & Publishing grant, University of Moratuwa, Sri
Lanka for publishing this paper.
REFERENCES
[1] Q. T. Ostrom, H. Gittleman, G. Truitt, A. Boscia, C. Kruchko, and
J. S. Barnholtz-Sloan, ‘‘CBTRUS statistical report: Primary brain and
other central nervous system tumors diagnosed in the United States in
2011–2015,’’ Neuro-Oncology, vol. 20, no. 4, pp. iv1–iv86, 2018.
[2] Q. T. Ostrom, H. Gittleman, J. Fulop, M. Liu, R. Blanda, C. Kromer,
Y. Wolinsky, C. Kruchko,and J. S. Barnholtz-Sloan, ‘‘CBTRUS statistical
report: Primary brain and central nervous system tumors diagnosed in the
united states in 2008-2012,’’ Neuro-Oncology, vol. 17, no. 4, pp. iv1–iv62,
Oct. 2015.
[3] Q. T. Ostrom, H. Gittleman, P. Liao, C. Rouse, Y. Chen, J. Dowling,
Y. Wolinsky, C. Kruchko, and J. Barnholtz-Sloan, ‘‘CBTRUS statisti-
cal report: Primary brain and central nervous system tumors diagnosed
in the United States in 2007–2011,’’ Neuro-Oncology, vol. 16, no. 4,
pp. iv1–iv63, Oct. 2014.
[4] M. Weller, W. Wick, K. Aldape, M. Brada, M. Berger, S. M. Pfister,
R. Nishikawa, M. Rosenthal, P. Y. Wen, R. Stupp, and G. Reifenberger,
‘‘Glioma,’’ Nature Rev. Disease Primers, vol. 1, no. 1, p. 15017, 2015.
[5] G. N. Fuller and B. W. Scheithauer, ‘‘The 2007 revised world health
organization (WHO) classification of tumours of the central nervous sys-
tem: Newly codified entities,’’ Brain Pathol., vol. 17, no. 3, pp. 304–307,
Jul. 2007.
[6] F. Sahm, D. Reuss, C. Koelsche, D. Capper, J. Schittenhelm, S. Heim,
D. T. W. Jones, S. M. Pfister, C. Herold-Mende, W. Wick, W. Mueller,
C. Hartmann, W. Paulus, and A. von Deimling, ‘‘Farewell to oligoastro-
cytoma: In situ molecular genetics favor classification as either oligo-
dendroglioma or astrocytoma,’’ Acta Neuropathol., vol. 128, no. 4,
pp. 551–559, 2014.
[7] K. K. Jain, ‘‘A critical overview of targeted therapies for glioblastoma,’’
Frontiers Oncol., vol. 8, p. 419, Oct. 2018.
[8] H. Ohgaki and P. Kleihues, ‘‘Epidemiology and etiology of gliomas,’’
Acta Neuropathol., vol. 109, no. 1, pp. 93–108, Jan. 2005.
[9] P. Yang, Y. Wang, X. Peng, G. You, W. Zhang, W. Yan, Z. Bao, Y. Wang,
X. Qiu, and T. and Jiang, ‘‘Management and survival rates in patients
with glioma in China (2004–2010): A retrospective study from a single-
institution,’’ J. Neuro-Oncol., vol. 113, no. 2, pp. 259–266, 2013.
[10] E. B. Claus, K. M. Walsh, J. K. Wiencke, A. M. Molinaro, J. L. Wiemels,
J. M. Schildkraut, M. L. Bondy, M. Berger, R. Jenkins, and M. Wrensch,
‘‘Survival and low-grade glioma: The emergence of genetic information,’’
Neurosurgical Focus, vol. 38, no. 1, p. E6, Jan. 2015.
[11] M. Klein, ‘‘Neurocognitive functioning in adult WHO grade II gliomas:
Impact of old and new treatment modalities,’’ Neuro-Oncology, vol. 14,
no. 4, pp. iv17–iv24, Sep. 2012.
[12] M. Moghtadaei, M. R. Hashemi Golpayegani, F. Almasganj, A. Etemadi,
M. R. Akbari, and R. Malekzadeh, ‘‘Predicting the risk of squamous dys-
plasia and esophageal squamous cell carcinoma using minimum classifi-
cation error method,’’ Comput. Biol. Med., vol. 45, pp. 51–57, Feb. 2014.
[13] M. S. Bal, V. K. Bodal, J. Kaur, M. Kaur, and S. Sharma, ‘‘Patterns
of cancer: A study of 500 Punjabi patients,’’ Asian Pacific J. Cancer
Prevention, vol. 16, no. 12, pp. 5107–5110, 2015.
[14] P. Sanghani, B. T. Ang, N. K. K. King, and H. Ren, ‘‘Overall survival
prediction in glioblastoma multiforme patients from volumetric, shape
and texture features using machine learning,’’ Surgical Oncol., vol. 27,
no. 4, pp. 709–714, Dec. 2018.
[15] M. Islam, V. J. M. Jose, and H. Ren, ‘‘Glioma prognosis: Segmenta-
tion of the tumor and survival prediction using shape, geometric and
clinical information,’’ in Brainlesion: Glioma, Multiple Sclerosis, Stroke
Traumatic Brain Injuries, A. Crimi, S. Bakas, H. Kuijf, F. Keyvan,
M. Reyes, and T. van Walsum, Eds. Cham, Switzerland: Springer, 2019,
pp. 142–153.
[16] U. Baid, S. U. Rane, S. Talbar, S. Gupta, M. H. Thakur, A. Moiyadi, and
A. Mahajan, ‘‘Overall survival prediction in glioblastoma with radiomic
features using machine learning,’’ Frontiers Comput. Neurosci., vol. 14,
no. 61, pp. 1–9, Aug. 2020, doi: 10.3389/fncom.2020.00061.
[17] N. Wijethilake, D. Meedeniya, C. Chitraranjan, and I. Perera, ‘‘Survival
prediction and risk estimation of glioma patients using mRNA expres-
sions,’’ in Proc. IEEE 20th Int. Conf. Bioinf. Bioeng. (BIBE), Oct. 2020,
pp. 35–42.
[18] P. Wang, Y. Li, and C. K. Reddy, ‘‘Machine learning for survival
analysis: A survey,’’ ACM Comput. Surv., vol. 51, no. 6, pp. 1–36,
2019.
[19] L. Ganau, M. Paris, G. K. Ligarotti, and M. Ganau, ‘‘Management of
gliomas: Overview of the latest technological advancements and related
behavioral drawbacks,’’ Behavioural Neurol., vol. 2015, pp. 1–7, 2015.
[20] S. Metcalfe and R. Grant, ‘‘Biopsy versus resection for malignant
glioma,’’ Cochrane Database Systematic Rev., vol. 1, no. 3, 2001,
Art. no. CD002034.
[21] N. Upadhyay and A. D. Waldman, ‘‘Conventional MRI evaluation of
gliomas,’’ Brit. J. Radiol., vol. 84, no. 2, pp. S107–S111, Dec. 2011.
[22] P. Kalavathi and V. S. Prasath, ‘‘Methods on skull stripping of MRI head
scan images—A review,’’ J. Digit. Imag., vol. 29, no. 3, pp. 365–379,
2016.
[23] A. Fedorov, R. Beichel, J. Kalpathy-Cramer, J. Finet, J.-C. Fillion-
Robin, S. Pujol, C. Bauer, D. Jennings, F. Fennessy, M. Sonka, J. Buatti,
S. Aylward, J. V. Miller, S. Pieper, and R. Kikinis, ‘‘3D slicer as an image
computing platform for the quantitative imaging network,’’ Magn. Reson.
Imag., vol. 30, no. 9, pp. 1323–1341, Nov. 2012.
[24] S. Bauer, T. Fejes, and M. Reyes, ‘‘A skull-stripping filter for ITK,’’
Insight J., vol. 2012, pp. 1–7, Jan. 2013.
[25] J. Lao, Y. Chen, Z.-C. Li, Q. Li, J. Zhang, J. Liu, and G. Zhai, ‘‘A deep
learning-based radiomics model for prediction of survival in glioblastoma
multiforme,’’ Sci. Rep., vol. 7, no. 1, pp. 1–8, Dec. 2017.
[26] N. Wijethilake, M. Islam, and H. Ren, ‘‘Radiogenomics model for overall
survival prediction of glioblastoma,’’ Med. Biol. Eng. Comput., vol. 58,
no. 8, pp. 1767–1777, Aug. 2020.
[27] N. Wijethilake, M. Islam, D. Meedeniya, C. Chitraranjan, I. Perera, and
H. Ren, ‘‘Radiogenomics of glioblastoma: Identification of radiomics
associated with molecular subtypes,’’ in Machine Learning in Clini-
cal Neuroimaging and Radiogenomics in Neuro-Oncology. Lima, Peru:
Springer, 2020, pp. 229–239.
[28] F. Tixier, H. Um, D. Bermudez, A. Iyer, A. Apte, M. S. Graham,
K. S. Nevel, J. O. Deasy, R. J. Young, and H. Veeraraghavan, ‘‘Preoper-
ative MRI-radiomics features improve prediction of survival in glioblas-
toma patients over MGMT methylation status alone,’’Oncotarget, vol. 10,
no. 6, p. 660, 2019.
[29] P. A. Yushkevich, J. Piven, H. C. Hazlett, R. G. Smith, S. Ho,
J. C. Gee, and G. Gerig, ‘‘User-guided 3D active contour segmentation of
anatomical structures: Significantly improved efficiency and reliability,’’
NeuroImage, vol. 31, no. 3, pp. 1116–1128, Jul. 2006.
[30] Y. Fu, N. Brown, S. Saeed, A. Casamitjana, Z. Baum, R. Delaunay,
Q. Yang, A. Grimwood, Z. Min, S. Blumberg, J. Iglesias,
D. Barratt, E. Bonmati, D. Alexander, M. Clarkson, T. Vercauteren,
and Y. Hu, ‘‘DeepReg: A deep learning toolkit for medical image
registration,’’ J. Open Source Softw., vol. 5, no. 55, p. 2705, Nov. 2020,
doi: 10.21105/joss.02705.
[31] B. B. Avants, N. Tustison, and G. Song, ‘‘Advanced normalization tools
(ANTS),’’ Insight J, vol. 2, no. 365, pp. 1–35, 2009.
[32] M. Jenkinson, C. F. Beckmann, T. E. Behrens, M. W. Woolrich, and
S. M. Smith, ‘‘FSL,’’ NeuroImage, vol. 62, no. 2, pp. 782–790, 2012.
[33] B. Fischl, ‘‘FreeSurfer,’’ NeuroImage, vol. 62, no. 2, pp. 774–781,
Aug. 2012.
[34] M. A. Hammoud, R. Sawaya, W. Shi, P. F. Thall, and N. E. Leeds,
‘‘Prognostic significance of preoperative MRI scans in glioblastoma mul-
tiforme,’’ J. Neuro-Oncol., vol. 27, no. 1, pp. 65–73, Jan. 1996.
[35] O. Ronneberger, P. Fischer, and T. Brox, ‘‘U-net: Convolutional networks
for biomedical image segmentation,’’ in Proc. Int. Conf. Med. Image
Comput. Comput.-Assist. Intervent. Munich, Germany: Springer, 2015,
pp. 234–241.
[36] I. D. Rubasinghe and D. A. Meedeniya, ‘‘Ultrasound nerve segmentation
using deep probabilistic programming,’’ J. ICT Res. Appl., vol. 13, no. 3,
pp. 241–256, 2019.
[37] J. Long, E. Shelhamer, and T. Darrell, ‘‘Fully convolutional networks
for semantic segmentation,’’ in Proc. IEEE Conf. Comput. Vis. Pattern
Recognit. (CVPR), Jun. 2015, pp. 3431–3440.
VOLUME 9, 2021 43187
N. Wijethilake et al.: Glioma Survival Analysis Empowered With Data Engineering—A Survey
[38] F. Isensee, P. Kickingereder, W. Wick, M. Bendszus, and K. H. Maier-
Hein, ‘‘Brain tumor segmentationand radiomics survival prediction: Con-
tribution to the brats 2017 challenge,’’ in Proc. Int. MICCAI Brainlesion
Workshop. Quebec City, QC, Canada: Springer, 2017, pp. 287–297.
[39] M. Islam, V. S. Vibashan, V. J. M. Jose, N. Wijethilake, U. Utkarsh, and
H. Ren, ‘‘Brain tumor segmentation and survival prediction using 3D
attention UNet,’’ in Brainlesion: Glioma, Multiple Sclerosis, Stroke Trau-
matic Brain Injuries, A. Crimi and S. Bakas, Eds. Cham, Switzerland:
Springer, 2020, pp. 262–272.
[40] L. Fidon, S. Ourselin, and T. Vercauteren, ‘‘Generalized wasserstein
dice score, distributionally robust deep learning, and ranger for brain
tumor segmentation: BraTS 2020 challenge,’’ 2020, arXiv:2011.01614.
[Online]. Available: http://arxiv.org/abs/2011.01614
[41] A. Myronenko, ‘‘3D MRI brain tumor segmentation using autoencoder
regularization,’’ in Proc. Int. MICCAI Brainlesion Workshop. Granada,
Spain: Springer, 2018, pp. 311–320.
[42] Z. Jiang, C. Ding, M. Liu, and D. Tao, ‘‘Two-stage cascaded U-net:
1st place solution to brats challenge 2019 segmentation task,’’ in Proc.
Int. MICCAI Brainlesion Workshop. Shenzhen, China: Springer, 2019,
pp. 231–241.
[43] S. Saman and S. Jamjala Narayanan, ‘‘Survey on brain tumor segmenta-
tion and feature extraction of MR images,’’ Int. J. Multimedia Inf. Retr.,
vol. 8, no. 2, pp. 79–99, Jun. 2019.
[44] L. Sun, S. Zhang, H. Chen, and L. Luo, ‘‘Brain tumor segmentation and
survival prediction using multimodal MRI scans with deep learning,’’
Frontiers Neurosci., vol. 13, no. 810, pp. 1–9, 2019.
[45] J. J. M. van Griethuysen, A. Fedorov, C. Parmar, A. Hosny, N. Aucoin,
V. Narayan, R. G. H. Beets-Tan, J.-C. Fillion-Robin, S. Pieper, and
H. J. W. L. Aerts, ‘‘Computational radiomics system to decode the
radiographic phenotype,’’ Cancer Res., vol. 77, no. 21, pp. e104–e107,
Nov. 2017.
[46] X. Feng, N. J. Tustison, S. H. Patel, and C. H. Meyer, ‘‘Brain tumor
segmentation using an ensemble of 3D U-Nets and overall survival pre-
diction using radiomic features,’’ Frontiers Comput. Neurosci., vol. 14,
p. 25, Apr. 2020.
[47] W. Han, L. Qin, C. Bay, X. Chen, K.-H. Yu, N. Miskin, A. Li, X. Xu,
and G. Young, ‘‘Deep transfer learning and radiomics feature prediction
of survival of patients with high-grade gliomas,’’ Amer. J. Neuroradiol.,
vol. 41, no. 1, pp. 40–48, Jan. 2020.
[48] K. Simonyan and A. Zisserman, ‘‘Very deep convolutional networks
for large-scale image recognition,’’ 2014, arXiv:1409.1556. [Online].
Available: http://arxiv.org/abs/1409.1556
[49] K. Chatfield, K. Simonyan, A. Vedaldi, and A. Zisserman, ‘‘Return of
the devil in the details: Delving deep into convolutional nets,’’ 2014,
arXiv:1405.3531. [Online]. Available: http://arxiv.org/abs/1405.3531
[50] G. Wang, W. Li, S. Ourselin, and T. Vercauteren, ‘‘Automatic brain tumor
segmentation using cascaded anisotropic convolutional neural networks,’’
in Proc. Int. MICCAI Brainlesion Workshop. Quebec City, QC, Canada:
Springer, 2017, pp. 178–190.
[51] Ö. Çiçek, A. Abdulkadir, S. S. Lienkamp, T. Brox, and O. Ronneberger,
‘‘3D U-Net: Learning dense volumetricsegmentation from sparse annota-
tion,’’ in Proc. Int. Conf. Med. Image Comput. Comput.-Assist. Intervent.
Athens, Greece: Springer, 2016, pp. 424–432.
[52] E. Puybareau, G. Tochon, J. Chazalon, and J. Fabrizio, ‘‘Segmentation
of gliomas and prediction of patient overall survival: A simple and fast
procedure,’’ in Proc. Int. MICCAI Brainlesion Workshop. Granada, Spain:
Springer, 2018, pp. 199–209.
[53] L. Chato, E. Chow, and S. Latifi, ‘‘Wavelet transform to improve accuracy
of a prediction model for overall survival time of brain tumor patients
based on MRI images,’’ in Proc. IEEE Int. Conf. Healthcare Informat.
(ICHI), Jun. 2018, pp. 441–442.
[54] Z. A. Shboul, M. Alam, L. Vidyaratne, L. Pei, M. I. Elbakary, and
K. M. Iftekharuddin, ‘‘Feature-guided deep radiomics for glioblas-
toma patient survival prediction,’’ Frontiers Neurosci., vol. 13, p. 966,
Sep. 2019.
[55] H. Dong, G. Yang, F. Liu, Y. Mo, and Y. Guo, ‘‘Automatic brain tumor
detection and segmentation using U-net based fully convolutional net-
works,’’ in Proc. Annu. Conf. Med. Image Understand. Anal. Edinburgh,
U.K.: Springer, 2017, pp. 506–517.
[56] D. Nie, H. Zhang, E. Adeli, L. Liu, and D. Shen, ‘‘3D deep learning
for multi-modal imaging-guided survival time prediction of brain tumor
patients,’’ in Proc. Int. Conf. Med. Image Comput. Comput.-Assist. Inter-
vent. Athens, Greece: Springer, 2016, pp. 212–220.
[57] M. Incoronato, M. Aiello, T. Infante, C. Cavaliere, A. Grimaldi,
P. Mirabelli, S. Monti, and M. Salvatore, ‘‘Radiogenomic analysis of
oncological data: A technical survey,’’ Int. J. Mol. Sci., vol. 18, no. 4,
p. 805, Apr. 2017.
[58] A. Welivita, I. Perera, D. Meedeniya, A. Wickramarachchi, and
V. Mallawaarachchi, ‘‘Managing complex workflows in bioinformatics:
An interactive toolkit with GPU acceleration,’’ IEEE Trans. Nanobiosci.,
vol. 17, no. 3, pp. 199–208, Jul. 2018.
[59] A. Alkuhlani, M. Nassef, and I. Farag, ‘‘Multistage feature selection
approach for high-dimensional cancer data,’’ Soft Comput., vol. 21,
no. 22, pp. 6895–6906, Nov. 2017.
[60] C. Hartmann, B. Hentschel, M. Tatagiba, J. Schramm, O. Schnell,
C. Seidel, R. Stein, G. Reifenberger, T. Pietsch, A. von Deimling,
M. Loeffler, and M. Weller, ‘‘Molecular markers in low-grade gliomas:
Predictive or prognostic?’’ Clin. Cancer Res., vol. 17, no. 13,
pp. 4588–4599, Jul. 2011.
[61] S. Zuo, X. Zhang, and L. Wang, ‘‘A RNA sequencing-based six-gene
signature for survival prediction in patients with glioblastoma,’’ Sci. Rep.,
vol. 9, no. 1, pp. 1–10, Dec. 2019.
[62] J. Xian, Q. Zhang, X. Guo, X. Liang, X. Liu, and Y. Feng, ‘‘A prognostic
signature based on three non-coding RNAs for prediction of the overall
survival of glioma patients,’’ FEBS Open Bio, vol. 9, no. 4, pp. 682–692,
2019.
[63] W.-J. Zeng, Y.-L. Yang, Z.-Z. Liu, Z.-P. Wen, Y.-H. Chen, X.-L. Hu,
Q. Cheng, J. Xiao, J. Zhao, and X.-P. Chen, ‘‘Integrative analysis of DNA
methylation and gene expression identify a three-gene signature for pre-
dicting prognosis in lower-grade gliomas,’’ Cellular Physiol. Biochem.,
vol. 47, no. 1, pp. 428–439, 2018.
[64] P. Mobadersany, S. Yousefi, M. Amgad, D. A. Gutman, J. S. Barnholtz-
Sloan, J. E. Velázquez Vega, D. J. Brat, and L. A. D. Cooper, ‘‘Predicting
cancer outcomes from histology and genomics using convolutional net-
works,’’ Proc. Nat. Acad. Sci. USA, vol. 115, no. 13, pp. E2970–E2979,
Mar. 2018.
[65] S. Rathore, M. Aksam Iftikhar, and Z. Mourelatos, ‘‘Prediction of overall
survival and molecular markers in gliomas via analysis of digital pathol-
ogy images using deep learning,’’ 2019, arXiv:1909.09124. [Online].
Available: http://arxiv.org/abs/1909.09124
[66] J. Zhao, L. Wang, D. Kong, G. Hu, and B. Wei, ‘‘Construction
of novel DNA methylation-based prognostic model to predict sur-
vival in glioblastoma,’’ J. Comput. Biol., vol. 27, no. 5, pp. 718–728,
May 2020.
[67] W.-Z. Gao, L.-M. Guo, T.-Q. Xu, Y.-H. Yin, and F. Jia, ‘‘Identification
of a multidimensional transcriptome signature for survival prediction of
postoperative glioblastoma multiforme patients,’’ J. Transl. Med., vol. 16,
no. 1, p. 368, Dec. 2018.
[68] H. Gittleman, A. E. Sloan, and J. S. Barnholtz-Sloan, ‘‘An independently
validated survival nomogram for lower-grade glioma,’’ Neuro-Oncology,
vol. 22, no. 5, pp. 665–674, May 2020.
[69] N. Beig, S. Singh, K. Bera, P. Prasanna, G. Singh, J. Chen, A. Saeed
Bamashmos, A. Barnett, K. Hunter, V. Statsevych, V. B. Hill, V. Varadan,
A. Madabhushi, M. S. Ahluwalia, and P. Tiwari, ‘‘Sexually dimorphic
radiogenomic models identify distinct imaging and biological pathways
that are prognostic of overall survival in glioblastoma,’’ Neuro-Oncology,
vol. 23, no. 2, pp. 251–263, Feb. 2021.
[70] R. J. Little and D. B. Rubin, Statistical Analysis With Missing Data,
vol. 793. Hoboken, NJ, USA: Wiley, 2019.
[71] L. Ladha and T. Deepa, ‘‘Feature selection methods and algorithms,’’ Int.
J. Comput. Sci. Eng., vol. 3, no. 5, pp. 1787–1797, 2011.
[72] L. Yu and H. Liu, ‘‘Feature selection for high-dimensional data: A fast
correlation-based filter solution,’’ in Proc. 20th Int. Conf. Mach. Learn.
(ICML), 2003, pp. 856–863.
[73] I. Guyon, J. Weston, S. Barnhill, and V. Vapnik, ‘‘Gene selection for can-
cer classification using support vector machines,’’ Mach. Learn., vol. 46,
nos. 1–3, pp. 389–422, 2002.
[74] I. T. Jolliffe and J. Cadima, ‘‘Principal component analysis: A review and
recent developments,’’ Phil. Trans. Roy. Soc. A, Math., Phys. Eng. Sci.,
vol. 374, no. 2065, Apr. 2016, Art. no. 20150202.
[75] A. J. Smola and B. Schölkopf, ‘‘A tutorial on support vector regression,’’
Statist. Comput., vol. 14, no. 3, pp. 199–222, Aug. 2004.
[76] W. Rasanjana, S. Rajapaksa, I. Perera, and D. Meedeniya, ‘‘A
SVM model for candidate Y-chromosome gene discovery in prostate
cancer,’’ in 11th Int. Conf. Bioinf. Comput. Biol. (BICOB) 2019,
pp. 129–138.
43188 VOLUME 9, 2021
N. Wijethilake et al.: Glioma Survival Analysis Empowered With Data Engineering—A Survey
[77] K. E. Emblem, M. C. Pinho, F. G. Zöllner, P. Due-Tonnessen,
J. K. Hald, L. R. Schad, T. R. Meling, O. Rapalino, and A. Bjornerud, ‘‘A
generic support vector machine model for preoperative glioma survival
associations,’’ Radiology, vol. 275, no. 1, pp. 228–234, Apr. 2015.
[78] L. Macyszyn, H. Akbari, J. M. Pisapia, X. Da, M. Attiah, V. Pigrish, Y. Bi,
S. Pal, R. V. Davuluri, L. Roccograndi, N. Dahmane, M. Martinez-Lage,
G. Biros, R. L. Wolf, M. Bilello, D. M. O’Rourke, and C. Davatzikos,
‘‘Imaging patterns predict patient survival and molecular subtype in
glioblastoma via machine learning techniques,’’ Neuro-Oncology, vol. 18,
no. 3, pp. 417–425, Mar. 2016.
[79] F. M. Khan and V. B. Zubek, ‘‘Supportvector regression for censored data
(SVRc): A novel tool for survival analysis,’’ in Proc. 8th IEEE Int. Conf.
Data Mining, Dec. 2008, pp. 863–868.
[80] E. Marubini, A. Morabito, and M. G. Valsecchi, ‘‘Prognostic factors and
risk groups: Some results given by using an algorithm suitable for cen-
sored survival data,’’ Statist. Med., vol. 2, no. 2, pp. 295–303, Apr. 1983.
[81] A. Ciampi, R. Bush, M. Gospodarowicz, and J. Till, ‘‘An approach to
classifying prognostic factors related to survival experience for non-
Hodgkin’s lymphoma patients: Based on a series of 982 patients: 1967–
1975,’’ Cancer, vol. 47, no. 3, pp. 621–627, 1981.
[82] L. Gordon and R. A. Olshen, ‘‘Tree-structured survival analysis,’’ Cancer
Treat. Rep., vol. 69, no. 10, pp. 1065–1069, 1985.
[83] A. Ciampi, J. Thiffault, J.-P. Nakache, and B. Asselain, ‘‘Stratification
by stepwise regression, correspondence analysis and recursive partition:
A comparison of three methods of analysis for survival data with covari-
ates,’’ Comput. Statist. Data Anal., vol. 4, no. 3, pp. 185–204, Oct. 1986.
[84] A. Ciampi, ‘‘Recursive partition and amalgamation (RECPAM) for cen-
sored survival data: Criteria for tree selection,’’ Stat. Softw. Newslett.,
vol. 14, no. 2, pp. 78–81, 1988.
[85] W. J. Curran, C. B. Scott, J. Horton, J. S. Nelson, A. S. Weinstein,
A. J. Fischbach, C. H. Chang, M. Rotman, S. O. Asbell, R. E. Krisch,
and D. F. Nelson, ‘‘Recursive partitioning analysis of prognostic factors
in three radiation therapy oncology group malignant glioma trials,’’ JNCI
J. Nat. Cancer Inst., vol. 85, no. 9, pp. 704–710, May 1993.
[86] G. Bauman, K. Lote, D. Larson, L. Stalpers, C. Leighton, B. Fisher,
W. Wara, D. Macdonald, L. Stitt, and J. G. Cairncross, ‘‘Pretreatment
factors predict overall survival for patients with low-grade glioma: A
recursive partitioning analysis,’’ Int. J. Radiat. Oncol. Biol. Phys., vol. 45,
no. 4, pp. 923–929, Nov. 1999.
[87] H. Jun Cho and S.-M. Hong, ‘‘Median regression tree for analysis of cen-
sored survival data,’’ IEEE Trans. Syst., Man, Cybern., A, Syst. Humans,
vol. 38, no. 3, pp. 715–726, May 2008.
[88] M. L. Gandía-González, S. Cerdán, L. Barrios, P. López-Larrubia,
P. G. Feijoó, A. Palpan, Jr., J. M. Roda, and J. Solivera, ‘‘Assessment of
overall survival in glioma patients as predicted by metabolomic criteria,’’
Frontiers Oncol., vol. 9, p. 328, May 2019.
[89] J. R. Quinlan, ‘‘Bagging, boosting, and C4. 5,’’ in Proc. AAAI/IAAI, vol. 1,
1996, pp. 725–730.
[90] L. Breiman, ‘‘Random forests,’’ Mach. Learn., vol. 45, no. 1, pp. 5–32,
2001.
[91] B. Efron, ‘‘Tibshirani,(1993), an introduction to the bootstrap,’’ in Mono-
graphs on Statistics and Applied Probability, vol. 1. Boca Raton, FL,
USA: CRC Press, 1960.
[92] P. Geurts, D. Ernst, and L. Wehenkel, ‘‘Extremely randomized trees,’’
Mach. Learn., vol. 63, no. 1, pp. 3–42, 2006.
[93] H. Ishwaran, U. B. Kogalur, E. H. Blackstone, and M. S. Lauer, ‘‘Random
survival forests,’’ Ann. Appl. Statist., vol. 2, no. 3, pp. 841–860, 2008.
[94] Y. S. Choi, S. S. Ahn, J. H. Chang, S.-G. Kang, E. H. Kim, S. H. Kim,
R. Jain, and S.-K. and Lee, ‘‘Machine learning and radiomic phenotyping
of lower grade gliomas: Improving survival prediction,’’ Eur. Radiol.,
vol. 30, pp. 3834–3842, Mar. 2020.
[95] Y. Freund and R. E. Schapire, ‘‘A decision-theoretic generalization of on-
line learning and an application to boosting,’’ in Proc. Eur. Conf. Comput.
Learn. Theory. Barcelona, Spain: Springer, 1995, pp. 23–37.
[96] J. H. Friedman, ‘‘Greedy function approximation: A gradient boosting
machine,’’ Ann. Statist., vol. 29, no. 5, pp. 1189–1232, 2001.
[97] T. Hothorn, ‘‘Survival ensembles,’’ Biostatistics, vol. 7, no. 3,
pp. 355–373, Dec. 2005.
[98] J. Zhu, J. Chen, W. Hu, and B. Zhang, ‘‘Big learning with Bayesian
methods,’’ Nat. Sci. Rev., vol. 4, no. 4, pp. 627–651, Jul. 2017.
[99] S. Rabinowicz, A. Hommersom, R. Butz, and M. Williams, ‘‘A prognostic
model of glioblastoma multiforme using survival Bayesian networks,’’
in Proc. Conf. Artif. Intell. Med. Eur. Vienna, Austria: Springer, 2017,
pp. 81–85.
[100] M. Zhou, B. Chaudhury, L. O. Hall, D. B. Goldgof, R. J. Gillies, and
R. A. Gatenby, ‘‘Identifying spatial imaging biomarkers of glioblastoma
multiforme for survival group prediction,’’ J. Magn. Reson. Imag., vol. 46,
no. 1, pp. 115–123, Jul. 2017.
[101] V. Bonato, V. Baladandayuthapani, B. M. Broom, E. P. Sulman,
K. D. Aldape, and K.-A. Do, ‘‘Bayesian ensemble methods for sur-
vival prediction in gene expression data,’’ Bioinformatics, vol. 27, no. 3,
pp. 359–367, Feb. 2011.
[102] S. P. Luttrell, ‘‘A Bayesian analysis of self-organizing maps,’’ Neural
Comput., vol. 6, no. 5, pp. 767–794, Sep. 1994.
[103] P. H. Abreu, M. S. Santos, M. H. Abreu, B. Andrade, and D. C. Silva,
‘‘Predicting breast cancer recurrence using machine learning techniques:
A systematic review,’’ ACM Comput. Surv., vol. 49, no. 3, pp. 1–40, 2016.
[104] S.-M. Lee and P. A. Abbott, ‘‘Bayesian networks for knowledge discovery
in large datasets: Basics for nurse researchers,’’ J. Biomed. Informat.,
vol. 36, nos. 4–5, pp. 389–399, Aug. 2003.
[105] N. Friedman, D. Geiger, and M. Goldszmidt, ‘‘Bayesian network classi-
fiers,’’ Mach. Learn., vol. 29, no. 2, pp. 131–163, Nov. 1997.
[106] M. Zhou, L. O. Hall, D. B. Goldgof, R. J. Gillies, and R. A. Gatenby, ‘‘Sur-
vival time prediction of patients with glioblastoma multiforme tumors
using spatial distance measurement,’’ Proc. SPIE, vol. 8670, Feb. 2013,
Art. no. 86702O.
[107] S. R. Piccolo and L. J. Frey, ‘‘Clinical and molecular models of glioblas-
toma multiforme survival,’’ Int. J. Data Mining Bioinf., vol. 7, no. 3,
p. 245, 2013.
[108] F. Rosenblatt, ‘‘The perceptron: A probabilistic model for information
storage and organization in the brain,’’ Psychol. Rev., vol.65, no. 6, p. 386,
1958.
[109] W. S. McCulloch and W. Pitts, ‘‘A logical calculus of the ideas immanent
in nervous activity,’’ Bull. Math. Biophys., vol. 5, no. 4, pp. 115–133,
Dec. 1943.
[110] D. Faraggi and R. Simon, ‘‘A neural network model for survival data,’’
Statist. Med., vol. 14, no. 1, pp. 73–82, Jan. 1995.
[111] T. Ching, X. Zhu, and L. X. Garmire, ‘‘Cox-nnet: An artificial neural
network method for prognosis prediction of high-throughput omics data,’’
PLOS Comput. Biol., vol. 14, no. 4, Apr. 2018, Art. no. e1006076.
[112] S. Yousefi, F. Amrollahi, M. Amgad, C. Dong, J. E. Lewis, C. Song,
D. A. Gutman, S. H. Halani, J. E. Velazquez Vega, D. J. Brat, and
L. A. D. Cooper, ‘‘Predicting clinical outcomes from large scale cancer
genomic profiles with deep survival models,’’ Sci. Rep., vol. 7, no. 1,
pp. 1–11, Dec. 2017.
[113] E. L. Kaplan and P. Meier, ‘‘Nonparametric estimation from incom-
plete observations,’’ J. Amer. Stat. Assoc., vol. 53, no. 282, pp. 457–481,
Jun. 1958.
[114] D. R. Cox, ‘‘Regression models and life-tables,’’ J. Roy. Stat. Soc., B
(Methodol.), vol. 34, no. 2, pp. 187–202, Jan. 1972.
[115] J. B.-K. Hsu, T.-H. Chang, G. A. Lee, T.-Y. Lee, and C.-Y. Chen,
‘‘Identification of potential biomarkers related to glioma survival by gene
expression profile analysis,’’ BMC Med. Genomics, vol. 11, no. S7, p. 34,
Mar. 2019.
[116] L. Wang, Z. Yan, X. He, C. Zhang, H. Yu, and Q. Lu, ‘‘A 5-gene prognos-
tic nomogram predicting survival probability of glioblastoma patients,’’
Brain Behav., vol. 9, no. 4, 2019, Art. no. e01258.
[117] T. Yang, P. Mao, X. Chen, X. Niu, G. Xu, X. Bai, and W. Xie,
‘‘Inflammatorybiomarkers in prognostic analysis for patients with glioma
and the establishment of a nomogram,’’ Oncol. Lett., vol. 17, no. 2,
pp. 2516–2522, Dec. 2018.
[118] D. Koshland Jr, K. Walsh, and D. LaPorte, ‘‘Sensitivity of metabolic
fluxes to covalent control,’’ in Current Topics in Cellular Regulation,
vol. 27. Amsterdam, The Netherlands: Elsevier, 1985, pp. 13–22.
[119] S. M. Lee, H.-J. Koh, D.-C. Park, B. J. Song, T.-L. Huh, and J.-W. Park,
‘‘Cytosolic NADP+-dependent isocitrate dehydrogenase status modu-
lates oxidative damage to cells,’’ Free Radical Biol. Med., vol. 32, no. 11,
pp. 1185–1196, 2002.
[120] A. L. Cohen, S. L. Holmen, and H. Colman, ‘‘IDH1 and IDH2 mutations
in gliomas,’’ Current Neurol. Neurosci. Rep., vol. 13, no. 5, pp. 765–773,
May 2013.
[121] C. Houillier, X. Wang, G. Kaloshi, K. Mokhtari, R. Guillevin, J. Laffaire,
S. Paris, B. Boisselier, A. Idbaih, F. Laigle-Donadey, K. Hoang-Xuan,
M. Sanson, and J.-Y. Delattre, ‘‘IDH1 or IDH2 mutations predict longer
survival and response to temozolomide in low-grade gliomas,’’ Neurol-
ogy, vol. 75, no. 17, pp. 1560–1566, Oct. 2010.
[122] B. Vogelstein, ‘‘P53 function and dysfunction,’’ Cell, vol. 70, no. 4,
pp. 523–526, Aug. 1992.
VOLUME 9, 2021 43189
N. Wijethilake et al.: Glioma Survival Analysis Empowered With Data Engineering—A Survey
[123] N. Ishii, D. Maier, A. Merlo, M. Tada, Y. Sawamura, A.-C. Diserens,
and E. G. Van Meir, ‘‘Frequent co-alterations of TP53, p16/CDKN 2A,
p14ARF , PTEN tumor suppressor genes in human glioma cell lines,’’
Brain Pathol., vol. 9, no. 3, pp. 469–479, 1999.
[124] Y. Okamoto, P.-L. Di Patre, C. Burkhard, S. Horstmann, B. Jourde,
M. Fahey,D. Schüler, N. M. Probst-Hensch, M. G. Yasargil,Y. Yonekawa,
U. M. Lütolf, P. Kleihues, and H. Ohgaki, ‘‘Population-based study on
incidence, survival rates, and genetic alterations of low-grade diffuse
astrocytomas and oligodendrogliomas,’’ Acta Neuropathol., vol. 108,
no. 1, pp. 49–56, Jul. 2004.
[125] X. Wang, J.-X. Chen, J.-P. Liu, C. You, Y.-H. Liu, and Q. Mao, ‘‘Gain
of function of mutant TP53 in glioblastoma: Prognosis and response to
temozolomide,’’ Ann. Surgical Oncol., vol. 21, no. 4, pp. 1337–1344,
Apr. 2014.
[126] T. T. Batchelor, R. A. Betensky, J. M. Esposito, L.-D.-D. Pham,
M. V. Dorfman, N. Piscatelli, S. Jhung, D. Rhee, and D. N. Louis, ‘‘Age-
dependent prognostic effects of genetic alterations in glioblastoma,’’ Clin.
Cancer Res., vol. 10, no. 1, pp. 228–233, Jan. 2004.
[127] K. Masutomi and W. C. Hahn, ‘‘Telomerase and tumorigenesis,’’ Cancer
Lett., vol. 194, no. 2, pp. 163–172, May 2003.
[128] M. Labussière, A. L. Di Stefano, V. Gleize, B. Boisselier, M. Giry,
S. Mangesius, A. Bruno, R. Paterra, Y. Marie, A. Rahimian,
G. Finocchiaro, R. S. Houlston, K. Hoang-Xuan, A. Idbaih,
J.-Y. Delattre, K. Mokhtari, and M. Sanson, ‘‘TERT promoter mutations
in gliomas, genetic associations and clinico-pathological correlations,’’
Brit. J. Cancer, vol. 111, no. 10, pp. 2024–2032, Nov. 2014.
[129] B. Heidenreich, P. S. Rachakonda, K. Hemminki, and R. Kumar, ‘‘TERT
promoter mutations in cancer development,’’ Current Opinion Genet.
Develop., vol. 24, pp. 30–37, Feb. 2014.
[130] P. J. Killela, Z. J. Reitman, and Y. Jiao, ‘‘TERT promoter mutations occur
frequently in gliomas and a subset of tumors derived from cells with
low rates of self-renewal,’’ Proc. Nat. Acad. Sci. USA, vol. 110, no. 15,
pp. 6021–6026, Apr. 2013.
[131] H. Arita, Y. Narita, S. Fukushima, K. Tateishi, Y. Matsushita, A. Yoshida,
Y. Miyakita, M. Ohno, V. P. Collins, N. Kawahara, S. Shibui, and
K. Ichimura, ‘‘Upregulating mutations in the TERT promoter commonly
occur in adult malignant gliomas and are strongly associated with total
1p19q loss,’’ Acta Neuropathol., vol. 126, no. 2, pp. 267–276, Aug. 2013.
[132] N. Nonoguchi, T. Ohta, J.-E. Oh, Y.-H. Kim, P. Kleihues, and H. Ohgaki,
‘‘TERT promoter mutations in primary and secondary glioblastomas,’’
Acta Neuropathol., vol. 126, no. 6, pp. 931–937, Dec. 2013.
[133] Y. Lee, J. Koh, S.-I. Kim, J. K. Won, C.-K. Park, S. H. Choi, and
S.-H. Park, ‘‘The frequency and prognostic effect of TERT promoter
mutation in diffuse gliomas,’’ Acta Neuropathol. Commun., vol. 5, no. 1,
p. 62, Dec. 2017.
[134] J. Liu, X. Zhang, X. Yan, M. Sun, Y. Fan, and Y. Huang, ‘‘Significance
of TERT and ATRX mutations in glioma,’’ Oncol. Lett., vol. 17, no. 1,
pp. 95–102, Oct. 2018.
[135] K. Gao, G. Li, Y. Qu, M. Wang, B. Cui, M. Ji, B. Shi, and P. Hou, ‘‘TERT
promoter mutations and long telomere length predict poor survival and
radiotherapy resistance in gliomas,’’ Oncotarget, vol. 7, no. 8, p. 8712,
2016.
[136] A. E. Pegg, M. E. Dolan, and R. C. Moschel, ‘‘Structure, function,
and inhibition of O6-alkylguanine-DNA alkyltransferase,’’ in Progress in
Nucleic Acid Research and Molecular Biology, vol. 51. Amsterdam, The
Netherlands: Elsevier, 1995, pp. 167–223.
[137] S. C. Schold, D. M. Kokkinakis, J. L. Rudy, R. C. Moschel, and
A. E. Pegg, ‘‘Treatment of human brain tumor xenografts with O6-
benzyl-2’-deoxyguanosine and BCNU,’’ Cancer Res., vol. 56, no. 9,
pp. 2076–2081, 1996.
[138] W. Wick, M. Weller, M. Van Den Bent, M. Sanson, M. Weiler, A. Von
Deimling, C. Plass, M. Hegi, M. Platten, and G. Reifenberger, ‘‘MGMT
testing—The challenges for biomarker-based glioma treatment,’’ Nature
Rev. Neurol., vol. 10, no. 7, p. 372, 2014.
[139] R. G. W. Verhaak, K. A. Hoadley, E. Purdom, V. Wang, and Y. Qi,
‘‘Integrated genomic analysis identifies clinically relevant subtypes of
glioblastoma characterized by abnormalities in PDGFRA, IDH1, EGFR,
and NF1,’’ Cancer Cell, vol. 17, no. 1, pp. 98–110, Jan. 2010.
[140] H. S. R. Rajula, G. Verlato, M. Manchia, N. Antonucci, and V. Fanos,
‘‘Comparison of conventional statistical methods with machine learning
in medicine: Diagnosis, drug development, and treatment,’’ Medicina,
vol. 56, no. 9, p. 455, Sep. 2020.
[141] S. Zhao, W.-P. Fung-Leung, A. Bittner, K. Ngo, and X. Liu, ‘‘Comparison
of RNA-seq and microarray in transcriptome profiling of activated T
cells,’’ PLoS ONE, vol. 9, no. 1, Jan. 2014, Art. no. e78644.
[142] Z. Bodalal, S. Trebeschi, T. D. L. Nguyen-Kim, W. Schats, and R. Beets-
Tan, ‘‘Radiogenomics: Bridging imaging and genomics,’’ Abdominal
Radiol., vol. 44, no. 6, pp. 1960–1984, Jun. 2019.
[143] F. Q. Wiki and Q. F.-P. W. Body, ‘‘The radiological society of North
America (RSNA) quantitative imaging biomarker alliance (QIBA) FDG-
PET technical committee: Proffered protocol for quantifying the standard
uptake (SUV) in patients with cancer,’’ no. Version 1.13, Tech. Rep.,
2016.
[144] M. M. Li, M. Datto, E. J. Duncavage, S. Kulkarni, N. I. Lindeman, S. Roy,
A. M. Tsimberidou, and C. L. Vnencak-Jones, ‘‘Standards and guidelines
for the interpretation and reporting of sequence variants in cancer: A joint
consensus recommendation of the association for molecular pathology,
American Society of clinical Oncology, and College of American pathol-
ogists,’’ J. Mol. Diagnostics, vol. 19, no. 1, pp. 4–23, 2017.
[145] S. M. Lundberg and S.-I. Lee, ‘‘A unified approach to interpreting
model predictions,’’ in Proc. Adv. Neural Inf. Process. Syst., I. Guyon,
U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan,
and R. Garnett, Eds. Red Hook, NY, USA: Curran Associates, 2017,
pp. 4765–4774. [Online]. Available: http://papers.nips.cc/paper/7062-a-
unified-approach-to-interpreting-model-predictions.pdf
[146] K. Simonyan, A. Vedaldi, and A. Zisserman, ‘‘Deep inside
convolutional networks: Visualising image classification models
and saliency maps,’’ 2013, arXiv:1312.6034. [Online]. Available:
http://arxiv.org/abs/1312.6034
[147] A. Welivita, I. Perera, and D. Meedeniya, ‘‘An interactive workflow
generator to support bioinformatics analysis through GPU acceleration,’’
in Proc. IEEE Int. Conf. Bioinf.Biomed. (BIBM), Nov. 2017, pp. 457–462.
[148] S. Rajapaksa, W. Rasanjana, I. Perera, and D. Meedeniya, ‘‘GPU acceler-
ated maximum likelihood analysis for phylogenetic inference,’’ in Proc.
8th Int. Conf. Softw. Comput. Appl., New York, NY, USA, Feb. 2019,
pp. 6–10.
[149] T. Karn, ‘‘High-throughput gene expression and mutation profiling:
Current methods and future perspectives,’’ Breast Care, vol. 8, no. 6,
pp. 401–406, 2013.
[150] M. J. Sheller, G. A. Reina, B. Edwards, J. Martin, and S. Bakas, ‘‘Multi-
institutional deep learning modeling without sharing patient data: A
feasibility study on brain tumor segmentation,’’ in Proc. Int. MICCAI
Brainlesion Workshop. Granada, Spain: Springer, 2018, pp. 92–104.
[151] M. J. Sheller, B. Edwards, G. A. Reina, J. Martin, S. Pati, A. Kotrotsou,
M. Milchenko, W. Xu, D. Marcus, R. R. Colen, and S. Bakas, ‘‘Federated
learning in medicine: Facilitating multi-institutional collaborations with-
out sharing patient data,’’ Sci. Rep., vol. 10, no. 1, pp. 1–12, Dec. 2020.
NAVODINI WIJETHILAKE (Student Member,
IEEE) received the B.Sc. degree in engineering
from the University of Moratuwa, Sri Lanka,
where she is currently pursuing the master’s
degree. She has worked as a Research Intern with
the Medical Mechatronics Laboratory, National
University of Singapore.
DULANI MEEDENIYA (Member, IEEE) received
the Ph.D. degree in computer science from the
University of St Andrews, U.K. She is currently
a Senior Lecturer with the Department of Com-
puter Science and Engineering, University of
Moratuwa, Sri Lanka. Her main research interests
include software modeling and design, bio-health
informatics, machine learning, data analytics and
technology enhanced learning. She is a Fellow
of HEA(UK), MIET, MIEEE, and a Charted
Engineer registered at EC (UK).
43190 VOLUME 9, 2021
N. Wijethilake et al.: Glioma Survival Analysis Empowered With Data Engineering—A Survey
CHARITH CHITRARANJAN received the B.Sc.
Eng. degree (Hons.) from the University of
Moratuwa, Sri Lanka, and the M.Sc. and Ph.D.
degrees from NDSU. He is currently a Lecturer
with the University of Moratuwa. His research
interests include data science biological, motor
traffic-related data, and MNBD.
INDIKA PERERA (Senior Member, IEEE)
received the B.Sc.Eng. (Hons.) and M.Sc. degrees
from the University of Moratuwa, Sri Lanka, and
the M.B.S. and P.G.D.B.M. degrees from the Uni-
versity of Colombo, and the Ph.D. degree from the
University of St Andrews, U.K. He is currently a
Senior Lecturer with the University of Moratuwa.
His research interests include research topics of
software architecture, software engineering; tech-
nology enhanced learning, UX and application
development for bio-health research. He is a Fellow of HEA, U.K., MIET,
SMIEEE, and a Charted Engineer registered at EC, U.K., and IE, Sri Lanka.
MOBARAKOL ISLAM (Member, IEEE) received
the Ph.D. degree from the National Univer-
sity of Singapore, Singapore, in 2019. He is
currently working with the Biomedical Image
Analysis Group, Imperial College London, U.K.
His current research interests include causal DL,
medical imaging, and image-guided surgery.
HONGLIANG REN (Senior Member, IEEE)
has navigated his academic journey through
The Chinese University of Hong Kong, Johns
Hopkins University, Children’s Hospital Boston,
Harvard Medical School, Children’s National
Medical Center, USA, and the National Univer-
sity of Singapore. His research interests include
biorobotics, intelligent control, medical mecha-
tronics, soft continuum robots, soft sensors, and
multisensory learning in medical robotics. He was
a recipient of the NUS Young Investigator Award and Engineering Young
Researcher Award, IAMBE Early Career Award 2018, Interstellar Early
Career Investigator Award 2018, and ICBHI Young Investigator Award 2019.
He currently serves as an Associate Editor for the IEEE TRANSACTIONS ON
AUTOMATION SCIENCE & ENGINEERING (T-ASE) and Medical & Biological
Engineering & Computing (MBEC).
VOLUME 9, 2021 43191
