A review on brain tumor diagnosis from MRI images: Practical implications, key achievements, and lessons learned

Abstract

The successful early diagnosis of brain tumors plays a major role in improving the treatment outcomes and thus improving patient survival. Manually evaluating the numerous magnetic resonance imaging (MRI) images produced routinely in the clinic is a difficult process. Thus, there is a crucial need for computer-aided methods with better accuracy for early tumor diagnosis. Computer-aided brain tumor diagnosis from MRI images consists of tumor detection, segmentation, and classification processes. Over the past few years, many studies have focused on traditional or classical machine learning techniques for brain tumor diagnosis. Recently, interest has developed in using deep learning techniques for diagnosing brain tumors with better accuracy and robustness. This study presents a comprehensive review of traditional machine learning techniques and evolving deep learning techniques for brain tumor diagnosis. This review paper identifies the key achievements reflected in the performance measurement metrics of the applied algorithms in the three diagnosis processes. In addition, this study discusses the key findings and draws attention to the lessons learned as a roadmap for future research.

Full text

Contents lists available at ScienceDirect
Magnetic Resonance Imaging
journal homepage: www.elsevier.com/locate/mri
Review Article
A review on brain tumor diagnosis from MRI images: Practical implications,
key achievements, and lessons learned
Mahmoud Khaled Abd-Ellah
a
, Ali Ismail Awad
b,c,*
, Ashraf A.M. Khalaf
d
, Hesham F.A. Hamed
d
a
Electronics and Communications Department, Al-Madina Higher Institute for Engineering and Technology, Giza, Egypt
b
Department of Computer Science, Electrical and Space Engineering, Luleå University of Technology, Luleå 97187, Sweden
c
Faculty of Engineering, Al-Azhar University, P.O. Box 83513, Qena, Egypt
d
Electronics and Communications Department, Faculty of Engineering, Minia University, Minia, Egypt
ARTICLE INFO
Keywords:
Brain tumor diagnosis
Computer-aided methods
MRI images
Tumor detection
Tumor segmentation
Tumor classification
Traditional machine learning techniques
Deep learning techniques
ABSTRACT
The successful early diagnosis of brain tumors plays a major role in improving the treatment outcomes and thus
improving patient survival. Manually evaluating the numerous magnetic resonance imaging (MRI) images
produced routinely in the clinic is a difficult process. Thus, there is a crucial need for computer-aided methods
with better accuracy for early tumor diagnosis. Computer-aided brain tumor diagnosis from MRI images consists
of tumor detection, segmentation, and classification processes. Over the past few years, many studies have
focused on traditional or classical machine learning techniques for brain tumor diagnosis. Recently, interest has
developed in using deep learning techniques for diagnosing brain tumors with better accuracy and robustness.
This study presents a comprehensive review of traditional machine learning techniques and evolving deep
learning techniques for brain tumor diagnosis. This review paper identifies the key achievements reflected in the
performance measurement metrics of the applied algorithms in the three diagnosis processes. In addition, this
study discusses the key findings and draws attention to the lessons learned as a roadmap for future research.
1. Introduction
Brain tumors consist of abnormally growing tissue resulting from
the uncontrolled multiplication of cells, and this tissue has no physio-
logical function inside the brain. Tumors not only increase the size of
and pressure in the brain but also cause swelling, all of which cause
abnormal neurological symptoms. According to the National Brain
Tumor Foundation (NBTF), the number of people in developed coun-
tries who die as a result of brain tumors has increased by 300% [1,2].
Brain tumors are classified as either metastatic or primary brain tumors.
In primary tumors, the cells are originally brain cells, but in metastatic
tumors, the cancer cells have spread into the brain from another in-
fected area of the body. Gliomas are the main type of tumor currently
attracting the interest of brain tumor researchers. The term glioma
describes different types of gliomas, ranging from high-grade (HG) tu-
mors, called glioblastoma multiforme (GBM), to low-grade (LG) tumors,
such as oligodendrogliomas or astrocytomas. Chemotherapy, radio-
therapy, and surgery may be applied to treat gliomas [3].
The main goal of computerized brain tumor diagnosis is to obtain
important clinical information regarding the tumor presence, location,
and type. The information obtained through clinical imaging can guide
and control any future interventions and thus leads to the correct di-
agnosis and treatment of the tumor. These automatic brain tumor di-
agnosis methods include different techniques that can be organized in a
pyramid. At each stage of the pyramid, distinct techniques are needed
to prepare, select, label, and describe the data, as indicated in [4].
Early tumor diagnosis plays a significant role in enhancing treat-
ment possibilities. Brain imaging techniques, such as positron emission
tomography (PET), single-photon emission computed tomography
(SPECT), computed tomography (CT), magnetic resonance imaging
(MRI), and magnetic resonance spectroscopy (MRS), are used to pro-
vide information about the location, size, shape, and type of brain
tumor to assist in the diagnosis. MRI provides rich information about
the anatomy of human tissues, and due to its widespread availability
and soft tissue contrast, it is considered to be a standard technique. MRI
uses radio frequency signals with a powerful magnetic field to produce
images of human tissues [5,6].
Brain tumor diagnosis consists of tumor detection, segmentation,
and classification processes. Brain tumor detection techniques are
mainly used to identify MRI images of tumors from a database, which is
https://doi.org/10.1016/j.mri.2019.05.028
Received 31 August 2018; Received in revised form 20 May 2019; Accepted 20 May 2019
*
Corresponding author.
E-mail addresses: mahmoudkhaled@ieee.org (M.K. Abd-Ellah), ali.awad@ltu.se (A.I. Awad), ashkhalaf@yahoo.com (A.A.M. Khalaf),
hfah66@yahoo.com (H.F.A. Hamed).
Magnetic Resonance Imaging 61 (2019) 300–318
0730-725X/ © 2019 Elsevier Inc. All rights reserved.
T

considered a basic and obvious process. However, brain tumor seg-
mentation techniques are used for localizing and isolating different
tumor tissues inside MRI images. Furthermore, brain tumor classifica-
tion techniques are used to classify abnormal images as malignant or
benign tumors. These three hybrid methods and techniques present
useful information to radiologists and aid in the understanding of MRI
information required for diagnosis.
Significant work in the field of brain tumor diagnosis has been
conducted by many researchers over the past several decades. Both
tumor segmentation and classification methods have been proposed.
The clinical acceptance of diagnosis methods has depended on the de-
gree of user supervision and the simplicity of computation [4]. How-
ever, the clinical applications are still limited, and although an ex-
tensive amount of work has been performed, clinicians still depend on
the manual projection of the tumor, probably because of the lack of a
connection between clinicians and researchers.
This study presents a review of the most important existing brain
tumor diagnosis methods. The survey focuses on MRI brain tumor di-
agnosis with traditional machine learning and deep learning techni-
ques. Although several reviews are available in the literature, a specific
focus is typically placed on one particular process, such as segmentation
in [4-9], classification in [10], or diagnosis in [3,11-13]. This article
offers a comprehensive overview of the whole brain tumor diagnosis
system in terms of tumor detection, segmentation, and classification. In
addition, the study covers the applications of traditional machine
learning and deep learning approaches in each phase or process of the
system.
The contributions of this study span multiple dimensions. First, the
study takes the entire brain tumor diagnosis system using MRI images
into account and considers the current conventional machine learning
and deep learning approaches for brain tumor diagnosis. Second, a
complete picture of the differences between and similarities of several
techniques is provided in terms of their performance in three brain
tumor diagnosis processes. Furthermore, the study introduces the
available MRI image databases that are used in evaluating the perfor-
mance of the reported techniques. Third, an extended discussion of the
current research findings and possible future improvements and trends
is provided. Fourth, a potential trial of integrating the three diagnosis
processes is outlined for future research using a single automated
system or model.
The remainder of this review is structured as follows. Section 2
presents necessary information on the brain tumor diagnosis frame-
work, the benefits of machine learning and deep learning techniques for
radiologists and clinicians, brain MRI images and available databases,
and the deep learning paradigm. Section 3 is dedicated to an in-depth
discussion of brain tumor diagnosis, which consists of tumor detection,
segmentation, and classification. This section introduces the key
achievements that have emerged from using classical machine learning
and trending deep learning techniques and related comparisons.
Section 4 is dedicated to discussing the research findings, the limita-
tions, and the lessons learned. Finally, conclusions are provided in
Section 5. A concise representation of the flow of this survey is shown in
Fig. 1.
2. Preliminaries
The general framework of a computer-aided diagnosis (CAD) system
for brain tumor diagnosis using MRI images, as summarized in Fig. 2,
consists of data collection, preprocessing, segmentation, feature ex-
traction, feature selection, feature reduction, classification, perfor-
mance evaluation and diagnosis. Data collection is the process of ob-
taining the brain images required for diagnosis that will be fed through
the diagnosis techniques. Brief explanations of the image types and the
databases are provided in Sections 2.2 and 2.3, respectively. The pre-
processing stage is a simple but necessary stage of brain image analysis.
Preprocessing is commonly used to improve the resolution and contrast
and to reduce the noise in the images. Several preprocessing approaches
can be used, such as unsharp masking, median filters and Wiener filters.
Median filters are most commonly used in the preprocessing stage to
preserve image edges [3].
Several segmentation methods are used to perform the segmenta-
tion task. One of these methodologies is the multilayer artificial neural
network (MANN). Many researchers have used the MANN and maintain
a similar structure and learning algorithm but mention it in different
ways, such as the feedforward backpropagation neural network
(FFBPNN), the backpropagation neural network (BPNN), and the mul-
tilayer perceptron (MLP). Other segmentation methodologies and
techniques include edge-based algorithms, region-based techniques,
and clustering algorithms (e.g., k-means, mean shift, fuzzy c-means
(FCM), and expectation maximization), and high performance can be
achieved by deep learning algorithms. There are different methods for
feature extraction, e.g., wavelet transform, texture features, Gabor
features, principal component analysis (PCA), decision boundary fea-
ture extraction, and spectral mixture analysis. Many different ap-
proaches have been developed for medical image segmentation and
analysis, as described in [14].
An increase in the feature vector dimension considerably decreases
the accuracy of the system. Hence, feature selection techniques are
applied to select the most important features. The popular feature se-
lection algorithms used in the literature are sequential backward se-
lection (SBS), genetic algorithm (GA), particle swarm optimization
(PSO), and sequential forward selection (SFS), while independent
component analysis (ICA), PCA, and kernel PCA are used for feature
dimensionality reduction [10].
Several classifiers are used in the detection and classification stages,
and they are reported here as they are mentioned in the published
papers, such as support vector machine (SVM), kernel SVM (KSVM),
feedforward backpropagation neural network (FFBPNN), self-orga-
nizing mapping neural network (SOMNN), backpropagation neural
network (BPNN), probabilistic neural network (PNN), artificial neural
network (ANN), probabilistic neural network-radial basis function
(PNN-RBF), normalized cross-correlation (NCC), PSO, sequential
minimal optimization (SMO), learning vector quantization (LVQ),
multilayer perceptron (MLP), k-nearest neighbor (KNN), hybrid of ge-
netic algorithm and support vector machine (GA-SVM), spectral clus-
tering independent component analysis (SC-ICA), least-squares feature
transformation (LSFT), fuzzy Hopfield neural network algorithm
(FHNN), sparse representation classification (SRC), unsupervised linear
discriminant analysis (ULDA), FCM and convolutional neural network
(CNN). Most of these tools are used efficiently; however, the highest
performances are achieved by deep learning algorithms.
2.1. Machine learning: a clinical perspective
From a clinical perspective, the goal of brain tumor diagnosis is to
accurately detect and localize tumor tissues from MRI images using
well-established clinical information and diagnostic features. The cor-
rect clinical diagnosis should lead to timely and appropriate disease
treatment. To achieve this goal, it is important to obtain clinical
knowledge and a database representing the information at a high level
from which a decision and diagnosis can be made [4]. Manual brain
tumor diagnosis is time consuming and less accurate due to the variety
of tumor shapes and types, as there are more than 120 known types of
brain tumors.
Machine learning has received considerable interest in modern
computing, and the medical field is one of the areas of interest. The
brain tumor diagnosis field has adopted different modern machine
learning techniques. For instance, advanced algorithms such as image
denoising [15,16], image reconstruction [17-19], skull stripping, and
registration [20] have been applied to simplify the use of brain images
and to enhance the obtained information. Therefore, machine learning
has created opportunities for collaboration among clinicians, engineers
M.K. Abd-Ellah, et al. Magnetic Resonance Imaging 61 (2019) 300–318
301