Conference PaperPDF Available

A Survey of Medical Image Classification Techniques

Authors:

Abstract

Medical informatics is the field that combines two medical data sources: biomedical record and imaging data. Medical image data is formed by pixels that correspond to a part of a physical object and produced by imaging modalities. Exploration of medical image data methods is a challenge in the sense of getting their insight value, analyzing and diagnosing of a specific disease. Image classification plays an important role in computer-aided-diagnosis and is a big challenge on image analysis tasks. This challenge related to the use of methods and techniques in exploiting image processing result, pattern recognition result and classification methods and subsequently validating the image classification result into medical expert knowledge. The main objective of medical images classification is not only to reach high accuracy but also to identify which parts of human body are infected by the disease. This paper reviewed the state-of-the-art of image classification techniques to diagnose human body disease. The review covered identification of medical image classification techniques, image modalities used, the dataset and trade off for each technique. At the end, the review showed the improvement of image classification techniques such as to increase accuracy and sensitivity value, and to be feasible employed for computer-aided-diagnosis are a big challenge and an open research.
978-1-5090-3352-2/16/$31.00 ©2016 IEEE 16-18 November 2016, Aston Tropicana Hotel, Bandung, Indonesia
2016 International Conference on Information Management and Technology (ICIMTech)
Page 56
A Survey of Medical Image Classification
Techniques
Eka Miranda, Mediana Aryuni
School of Information Systems
Bina Nusantara University
Jakarta, Indonesia
ekamiranda@binus.ac.id, mediana.aryuni@binus.ac.id
E. Irwansyah
School of Computer Science
Bina Nusantara University
Jakarta, Indonesia
eirwansyah@binus.edu
Abstract— Medical informatics is the study that combines two
medical data sources: biomedical record and imaging data.
Medical image data is formed by pixels that correspond to a part
of a physical object and produced by imaging modalities.
Exploration of medical image data methods is a challenge in the
sense of getting their insight value, analyzing and diagnosing of a
specific disease. Image classification plays an important role in
computer-aided-diagnosis and is a big challenge on image
analysis tasks. This challenge related to the usage of methods and
techniques in exploiting image processing result, pattern
recognition result and classification methods and subsequently
validating the image classification result into medical expert
knowledge. The main objective of medical images classification is
not only to reach high accuracy but also to identify which parts of
human body are infected by the disease. This paper reviewed the
state-of-the-art of image classification techniques to diagnose
human body disease. The review covered identification of
medical image classification techniques, image modalities used,
the dataset and trade off for each technique. At the end, the
reviews showed the improvement of image classification
techniques such as to increase accuracy and sensitivity value and
to be feasible employed for computer-aided-diagnosis are a big
challenge and an open research.
Keywords—medical informatics; image classification; disease
diagnosis
I.
I
NTRODUCTION
Medical Informatics is the study that intersects Information
Technology and healthcare [1]. In the implementation of field
research, medical informatics combines two medical data
sources. They are the biomedical record and imaging data,
which have specific characteristic [2]. Digital image data is
formed by pixels that correspond to a part of a physical object
as a result of imaging modalities [2]. In contrast biomedical
record is formed by record from patient medical tests. The
difference characteristic between biomedical record and
imaging data makes difference methodologies and techniques
needed to explore them. Researchers and methods to explore
biomedical record have already exploited [3, 4].
Medical image data is produced by imaging modalities. The
issue of this field is how to extract the image and classify the
extraction result into the similar pattern then identify and
understand which parts of human body are affected by the
specific disease from image classification result [5].
There are three stages of Medical Image analysis tasks
include (1) feature extraction and representation, (2) feature
selection that will be used for classification, and (3) feature and
image classification [6]. Moreover, computer-aided-diagnosis
needs the important role of image classification. Furthermore
medical image classification has three main steps: pre-
processing, feature extraction and classification [7]. After pre-
processing step then it needs to extract features of interest part
from the image for further analysis. The purpose of the pattern
classification system is to map the input variables (such as
record data or image data) become the output variables (to
represent one specific class (with a disease or with no a disease
class) [6].
Image classification is a big challenge on image analysis
tasks, especially the selection of methods and techniques in
exploiting the result of image processing and pattern
recognition, classification methods, subsequently validating the
image classification result into medical expert knowledge [7].
The main objective of medical images classification is not only
to reach high accuracy but also to identify which parts of
human body are infected by the disease. [8]. In the future, an
automatic diagnosis technique with image data is needed to be
developed for better clinical care [7].
Since image classification research is still an open area and
big challenge to validate the image classification result into
medical expert knowledge, this paper focuses only on detailed
review of mining medical image classification technique with
the state-of-the-art that addressing this issue. This paper will
not review about feature extraction and feature selection. The
aim of the review is to give the reader wide-ranging review of
mining medical image classification techniques includes their
pros and cons.
II. I
MAGING
M
ODALITIES
The source of medical images data is generated from
Biomedical Devices, which use the imaging techniques like
Computed Tomography (CT), Magnetic Resonance Imaging
(MRI) and mammogram [9].
There are several medical imaging modalities that involve
ionizing radiation, nuclear medicine, magnetic resonance,
ultrasound, and optical methods as a modality media. Each
modality media has a special characteristic and differences
Authorized licensed use limited to: Universitas Indonesia. Downloaded on April 28,2023 at 09:24:02 UTC from IEEE Xplore. Restrictions apply.
978-1-5090-3352-2/16/$31.00 ©2016 IEEE 16-18 November 2016, Aston Tropicana Hotel, Bandung, Indonesia
2016 International Conference on Information Management and Technology (ICIMTech)
Page 57
response to human body structure and organs tissue [5]. There
are four imaging modalities [9]:
A. Projectional Imaging
X-rays are a form of electromagnetic radiation (EM), which
has a wavelength range between 0.1-10 nm. They are translated
into photons with energy levels, 12-125 keV. The x-ray
imaging projection used almost at the same time with the need
to use laboratory testing as a medical diagnostic tool. Image
formation process is divided into three main steps: Image pre-
read, Image main read, Image processing [9].
B. Computed Tomography (CT)
The conventional x-ray imaging projection sometimes fails
in achieving good results because of tiny differences in
attenuation (less than 5%). CT improves the subject contrast
using discrimination less than 1%. The application for cancer
screening such as lung and virtual colonoscopy often uses CT.
There are several variations of CT imaging, namely: Positron
emission tomography (PET) / CT, CT perfusion, CT
angiography, Dual source and dual energy CT [9].
C. Magnetic Resonance (MR)
A powerful magnetic field is used in Magnetic Resonance
Imaging method (MR) for the nuclear magnetization
alignment of hydrogen atoms in water molecules. MR became
the standard of cross-sectional imaging modalities that useful
to visualize soft tissues (such as muscle, brain), fat and bone
(especially marrow bone) [9].
D. Ultrasound Imaging
The high- sound waves with the frequency range from 1-
20 MHz that can be applied to produce cross-sectional images
of the human body. The strength of the echo ultrasound return
depends on the characteristics of biological tissue which they
pass through.
III. M
EDICAL
I
MAGE
A
NALYSIS
T
ASKS
An image is considered as a representation of an object
with specific properties that are employed in image processing.
The medical image analysis tasks consist of Feature extraction
and representation, Feature selection that will be used for
classification, and Feature and image classification.
A. Feature Extraction and Representation
Features are an important measurement for image
understanding, especially the feature representation of the
segmented region that used for object classification and
analysis [10]. The techniques for Feature extraction and
representation include:
Statistical Pixel-Level (SPL) Features
These features provide quantitative information about
the pixels within a segmented region. The SPL
features include: mean, variance, and a histogram of
the gray values of pixels in the region, and
additionally: the area of the region and information
about the contrast of pixels within the region and
edge gradient of boundary pixels [10].
Shape Feature
These features provide information about shape
characteristic of the region boundary, include
circularity, compactness, moments, chain-codes, and
Hough transform. Morphological processing methods
have also been used for shape description [10].
Texture Features
These features provide information about the local
texture within the region or related area of the image,
which are calculated using the second-order
statistical histogram or co-occurrence matrices.
Besides wavelet processing method for spatial-
frequency analysis is employed for local texture
information representation [10].
Relational Features
These features provide information about the
relational and hierarchical structure of the regions
related to a single object or a group of objects [10].
B. Feature Selection for Classification
Feature selection used to discover the important features
that most appropriate for the classification task. Selection of
correlated features for dimensionality reduction in the
classification task can improve the computational efficiency
and classification performance. The final set of features can be
determined through data correlation, clustering, and analysis
algorithms to explore similarity patterns in the training data
[10]. A typical classification system showed in Fig.1.
Fig. 1. A classification system [6]
Feature selection for classification techniques are:
Linear Discriminant Analysis (LDA)
The purpose of LDA method is to discovery a linear
combination of features which able to give the best
possible separation among different classes of data in
the feature space. It can reduce dimensionality space
for classification and also give better classification
accuracy [10].
Principal Component Analysis (PCA)
It is an efficient method of dimensionality reduction
of a data set with a big number of interrelated
variables. However, for data with sparse distribution
and noise, PCA method may not provide optimal
selection of features [10].
GA (Genetic Algorithms)-Based Optimization
It is a robust technique for search optimization that
uses natural selection principles. It utilizes prior
information and using selection for survival, and is
able to adapt to the specific parameter issues. The
parameters are encoded as binary strings that are
associated with the fitness measurement [10].
Authorized licensed use limited to: Universitas Indonesia. Downloaded on April 28,2023 at 09:24:02 UTC from IEEE Xplore. Restrictions apply.
978-1-5090-3352-2/16/$31.00 ©2016 IEEE 16-18 November 2016, Aston Tropicana Hotel, Bandung, Indonesia
2016 International Conference on Information Management and Technology (ICIMTech)
Page 58
TABLE I. C
OMPILATION OF MEDICAL IMAGE CLASSIFICATION TECHNIQUES
Author Name Year Methodolog
y
Pros and Cons
Method Imaging Modalities
A. A. A. Setio et al
[11]
2016 Multi-View Convolutional
Networks (ConvNets)
Pulmonary CT False positive reduction.
The CAD sensitivity performance should be enhanced.
D. Mittal and A. Rani
[12]
2016 SVM Ultrasound image High accuracy.
Each classifier being trained on only two out of N classes.
G. Van Tulder and M.
De Bruijne [13]
2016 Convolutional classification
Restricted Boltzmann Machine
(RBM).
Lung CT High mean classification accuracy.
Suitable for smaller representations learning with smaller
filters or hidden nodes.
J. Hong et al [14]
2016 Principal Nested Spheres
(PNS), Distance Weighted
Discrimination(DWD)
MRI AUC > 0.600.
Apply PNS separately.
K. Seetharaman and S.
Sathiamoorthy [15]
2016 Adaptive Binary T
r
ee Based
Support Vector Machine
(ABTSVM)
CT, MRI, Microscopy,
Mammogram,
Ultrasound, X-ray and
Endoscopy images
Low computational and storage cost.
The relevance judgments are performed using the ground
truth and the subjectivity of the individual user.
K. Sirinukunwattana et
al [16]
2016 Neighboring Ensemble
Predictor (NEP) +
Convolutional Neural
Network (CNN)
Histopathology images Accurately predict.
The Weighted Average F1 score and Multiclass AUC
result not considerably different with softmax CNN +
SSPP.
M. Anthimopoulos et
al [17]
2016 Convolutional
Neural Network (CNN)
Lung CT Scan
Drawback
High classification performance.
The training time becomes slower due to very large
number of parameters.
M. J. J. P. Van
Grinsven et al [18]
2016 Convolutional neural
networks (CNNs) + Selective
Sampling (SeS)
Color fundus image
High performance.
Uses the reference guide from a single expert.
Q. Dou et al [19]
2016 3D Convolutional Neural
Network
(CNN)
Cerebral micro-
b
leeds
(CMBs) MRI
High sensitivity 93:16%.
The accuracy and detection speed are not balance.
A. Masood and A. Al-
jumaily [20]
2015 SVM Biopsy samples High accuracy.
The error rate of classification decreased about 16.5% for
Histopathological and 6% for Dermoscopic images.
F. Khalvati, A. Wong,
and M. A. Haider [21]
2015 SVM classifier.
Multi-parametric
magnetic resonance
imaging (MP-MRI)
High sensitivity and specificity (>80%).
A limited number of datasets and the target of Gleason
score is >= 7, the proposed model was not assessed by
clinicians.
K. Chung et al [22]
2015 Pre-Trained Convolutional
Neural Networks (CNN)
CT scan AUC = 0.868.
Time-consuming since Peri-Fissural Nodules (PFN)
characterization was subjective, it suggests the increment
of the number of 2D views may give the higher accuracy
of characterization.
V. Gopalakrishnan, P.
G. Menon, and S.
Madan [23]
2015 Bayesian rule
learning (BRL) methods
Cardiovascular
Magnetic Resonance
Imaging (cMRI)
High accuracy.
A limited number of datasets.
Y. Iwahori et al [24] 2015
K
-means++ Endoscope
The accuracy is higher because using the edge-
b
ased
features.
The computational time was decreased if HOG features
used to detect the polyp region.
Y. Song et al [25]
2015 Locality-constrained Sub-
cluster Representation
Ensemble (LSRE)
High Resolution
Computed
Tomography (HRCT)
High accuracy.
The Locality-constrained Linear Coding (LLC) did not
use advanced distance function.
B. Manju, K.
Meenakshy, and R.
Gopikakumari [26]
2014 KNN classifie
r
CT images High accuracy.
A limited number of datasets.
G. N. Balaji, T. S.
Subashini, and N.
Chidambaram [27]
2014 BPNN, SVM Echocardiogram High performance (87.5%).
Exclude other views include subcostal and Doppler view.
L. Ai, X. Gao, and J.
Xiong [28]
2014 Mean-Shift Clustering
(MSC)
Functional magnetic
resonance imaging
(fMRI)
Low false positive rate.
The significance levels could not be easily theoretically
measured. Still, it may show some challenges when it
needs very accurate comparisons.
S. Yazdani, R. Yusof,
A. Riazi, and A.
Karimian [29]
2014 SVM Magnetic Resonance
Images (MRI)
Desirable performance.
Not consider sub-cortical structures and 3 T images.
Reduce error rate from 30% down to less than 10%. .
L. Tan et al [30] 2013 SVM classifiers for sMRI,
SVM classifiers and
Hierarchical clustering for
fMRI.
The structural MR
images (sMRI) and
functional MR images
(fMRI)
High accuracy.
Not include tissue density maps and functional
connectivity networks. fMRI classifier had difficulty in
classifying some of the negative subjects.
R. To
m
ari et al [31]
2014 ANN Light microscope that
equipped with
DinoEye Eyepiece
High accuracy.
A limited number of datasets.
Authorized licensed use limited to: Universitas Indonesia. Downloaded on April 28,2023 at 09:24:02 UTC from IEEE Xplore. Restrictions apply.
978-1-5090-3352-2/16/$31.00 ©2016 IEEE 16-18 November 2016, Aston Tropicana Hotel, Bandung, Indonesia
2016 International Conference on Information Management and Technology (ICIMTech)
Page 59
Camera
A. Andrade et al [32]
2012 Compare ANN, SVM and
kNN classifier.
Ultrasound
Images
Higher accuracy for SVM classifier.
A limited number of datasets.
J. Miguel, P. Dias, C.
Manta, A. Luís, and S.
Cruz [33]
2012 Feed-Forward
Backpropagation Neural
Network
Digital Retinal Images
High sensitivity and specificity.
The classification is not significantly affected by the
randomized initialization of the neural networks weights.
M. A. Dabbah et al
[34]
2011 The multi-scale
dual-model using the NNT
pixel classification
Corneal confocal
microscopy (CCM)
image
Maximum sensitivity and specificity is the EER of
15.44%.
Compared to manual analysis delivers equivalent results.
E. J. Ciaccio et al [35] 2010 Classification method (using
threshold
or incremental learning)
Capsule endoscopy
video clip images
High threshold classifier and high incremental classifier
(high sensitivity and specificity).
H. Wang and B. Fei
[36]
2009 Multiscale fuzzy C-means
(MsFCM) classification
method
MR images Overlap ratio is greater than 90% and valida
t
ed by the
ground truth.
Classification method was not sensitive to the threshold
etween 0.8 and 0.9.
M. Mete, L. Hennings,
H. J. Spencer, and U.
Topaloglu [37]
2009 Support Vecto
r
Machines
Regular digital camera
which is attached to a
microscope
Achieving 0.90 for overall f-measure.
Only measures set-based dimensional fraction component.
Sub-image has the same opportunities to be categorized as
the positive or negative value.
R. Marée, P. Geurts,
and L. Wehenkel [38]
2007 Random sub windows and
extremely randomized trees
b
iology cell image The accuracy is good enough without a specific pre-
processing neither the domain knowledge.
The issue is misclassification error rates
M. Niemeijer, M. D.
Abra, and B. Van
Ginneken [39]
2006 Suppor
t
Vector Machine
Retinopathy screening The ROC curve is 0.9968 and the accuracy is 0.974.
The attained accuracy of 0.974 is still can be improved.
S. P. Awate, T.
Tasdizen, N. Foster,
and R. T. Whitaker
[40]
2006 Markov statistics non-
parametrically
Magnetic Resonance
(MR) images
Higher mean (by a couple of percents) and lower standard
deviation
Only use one algorithm "nearest neighbors" in Markov
neighborhood technique
C. Feature and Image Classification
The selected features of image representation that are
generated from feature selection, are used in object
recognition and characterization. In the medical imaging
analysis, features and measurements can also be used for
region segmentation to extract meaningful structures,
subsequently, interpret the result using knowledge-based
model and classification methods [6]. Feature and image
classification techniques namely:
Statistical Classification Methods
The categories of these methods are an unsupervised
and supervised approach. The unsupervised methods
cluster the data based on their separation in the feature
space, include K-means and fuzzy clustering. On the
other hand, a supervised approach needs training data,
test data, and class label to classify the data, include
probabilistic methods like the nearest neighbor and
Bayesian classifier [6].
Rule-Based Systems
The system analyzes the feature vector using multiple
sets of rules that are designed to test specific
conditions in the feature vector database to set off an
action. The rules consist of two parts: condition
premises and actions, which are generated based on an
expert knowledge to deduce the action when the
conditions are satisfied. The action which part of the
rule could change the database state or label of a
feature vector based on a specific state of analysis.
Usually, a rule-based system consists of three sets of
rules: supervisory or strategy rules, a focus of
attention rules, and knowledge rules. The supervisory
or strategy rules control the analysis process and
provide the control actions include starting and
stopping action. The strategy rules determine which
rules would be tested during the analysis process. The
focus-of-attention rules provide specific features
within analysis process by accessing and extracting
the information or features from the database.
Subsequently, the rules convey the information from
the input (database) into the activity center where the
implementations of knowledge rules are scheduled.
Finally, the knowledge rules analyze the information
related to the required conditions then execute an
action that changes the output database [6].
Neural Network Classifiers
Artificial neural network paradigms for feature
classification, object recognition and image
interpretation namely back-propagation, radial basis
function, associative memories, and self-organizing
feature maps. At that time fuzzy system-based
approaches have been applied in artificial neural
networks for better classification and generalization
result [6].
Support Vector Machine (SVM) for Classification
The Relevance Vector Machine (RVM) combines
regression, classification, and a Bayesian probabilistic
principle. The other models for pattern classification
utilizing theoretical approaches include kernel-based
classifier and linear programming perturbation-based
methods [6].
Compilation of medical image classification techniques
showed in TABLE I.
Authorized licensed use limited to: Universitas Indonesia. Downloaded on April 28,2023 at 09:24:02 UTC from IEEE Xplore. Restrictions apply.
978-1-5090-3352-2/16/$31.00 ©2016 IEEE 16-18 November 2016, Aston Tropicana Hotel, Bandung, Indonesia
2016 International Conference on Information Management and Technology (ICIMTech)
Page 60
IV. D
ISCUSSION
There are several challenges for the computer to
understand an image. First: diversity and number of image
data increase continuously, Second: mathematical and
statistical formulation and techniques, and how they could be
adjusted in the medical domain, Third: computing power [2].
Subsequently, the challenge of medical image analysis
model development comes from: (1) Complexity of the image
data: Although the quality of an image generated by medical
imaging modalities could be improved, the information
captured by medical imaging modalities not always complete
and clear. This problem due to lack of technology and imaging
process which produce incompetent resolution and artifacts
[2], (2) Complexity of the model or prototype: Medical
imaging analysis involves various human body structures,
features on the different application. The objects of medical
imaging analysis could be atomic structures (e.g., spine, brain
cortex, coronary arteries), pathological tissue (e.g., tumor,
myocardial infarction, inflammation, edema), functional areas
(e.g., motor cortex, glucose metabolism), or artificial objects
(e.g., implants, electrodes, catheter tips) which show important
biological variability of a subject. Medical imaging model or
prototype explains the object based on previous knowledge
that must identify diverse human body visualization [2], (3)
Validation: A common problem in medical imaging analysis is
data validation [2]. Human measurements are not absolutely
accurate. Medical imaging algorithm requires a ground truth
standard to validate their result [41-42].
TABLE II shows classification scheme of medical image
classification for each image classification technique. Neural
network classifier and SVM are the most used technique for
image classification and they could classify image from
almost all image modalities, additionally, many researchers
that used this technique showed high accuracy and sensitivity,
reasonable prediction and high classification performance
result [12, 16-21, 26-27, 29-34]. In addition, high sensitivity
and the specificity value of the research should be completed
with a suitable number of datasets with the purpose of the
feasibility to be employed for computer-aided-diagnosis.
Furthermore, many medical image analysis researchers used
image data from CT, MR and ultrasound imaging modalities
[11-15,17,19,21-23,25,26,28-30,32,36,40], since these three
imaging modalities could be used to determine the presence or
absence of the lesion based on a patient history [9].
Additionally, Magnetic Resonance Imaging does not use x-
rays [9]. For the future work, the neural network plays an
important role in classification since it can be used with its
supervised and unsupervised techniques [16]. Additionally,
fuzzy system-based approaches have been applied in the
artificial neural network for better classification result [6, 36].
Another interesting challenge is reducing an error rate by
applying hybrid approach since applying single approach still
give a high error rate result (above 10%) [20, 29, 30, 37].
TABLE II. C
LASSIFICATION
S
CHEME OF MEDICAL IMAGE
CLASSIFICATION TECHNIQUE
Image Classification Techniques Imaging Modalities
Statistical Classification Methods X-Ray CT MR Ultrasound
V. Gopalakrishnan, P. G. Menon,
and S. Madan [23]
Y. Iwahori et al [24]
H. Wang and B. Fei [36]
S. P. Awate, T. Tasdizen, N. Foster,
and R. T. Whitaker [40]
Neural Network Classifiers
M. Anthimopoulos et al [17]
M. J. J. P. Van Grinsven et al [18]
Q. Dou et al [19]
Support Vector Machine (SVM)
D. Mittal and A. Rani [12]
A. Masood and A. Al-jumaily [20]
F. Khalvati, A. Wong, and M. A.
Haider [21]
K. Chung et al [22]
B. Manju, K. Meenakshy, and R.
Gopikakumari [26]
S. Yazdani, R. Yusof, A. Riazi, and
A. Karimian [29]
M. Niemeijer, M. D. Abra, and B.
Van Ginneken [39]
V. CONCLUSION
AND
FUTURE
WORK
Medical image classification is an interesting research area,
it combines the diagnosis problem and analysis purposes in the
medical field. This paper has provided the detailed review of
image classification techniques for diagnosis of human body
disease include imaging modalities used, each dataset and pros
and cons for each technique. For the future work, the
improvement of image classification techniques will increase
accuracy value and subsequently feasible to be employed for
computer-aided-diagnosis, and more robust methods are being
developed.
A
CKNOWLEDGMENT
We gratefully acknowledge the supporting from institutions
that support this paper. Thank you to Research and Technology
Transfer office (RTTO), Bina Nusantara University which
helped the internal funding for this paper.
R
EFERENCES
[1] R. Procter, “Definition of Health Informatics”. Internet:
https://www.nlm.nih.gov/hsrinfo/informatics.html, Aug. 01, 2016, [Aug.
03, 2016].
[2] A. A. T. Bui, R. K. Taira, and H. Kangarloo, 2010,”Introduction,” in
Medical Imaging Informatics. New York: Springer, 2010, pp 3-14.
[3] F. S. Roque, P. B. Jensen, H. Schmock, M. Dalgaard, M. Andreatta, T.
Hansen, K. Søeby, S. Bredkjær, A. Juul, T. Werge, and L. J. Jensen,
“Using Electronic Patient Records to Discover Disease Correlations and
Stratify Patient Cohorts,” PLOS Computational Technology, vol. 7, no.
8, pp. 1-10, 2011.
[4] Y. Qiu, S. Zhen, M. Zhou, and L. Li, “Continuously improve the
medical care quality and hospital management level through medical
information system construction,” Journal of Translational Medicine,
vol. 10, suppl 2, p. A56, 2012.
[5] D. Aberle, S. El-Saden, P. Abbona, A. Gomez, K. Motamedi, N.
Ragavendra, L. Bassett, L. Seeger, M. Brown, K. Brown, A. A. T. Bui,
and H. Kangarloo, “A Primer on Imaging Anatomy and Physiology”, in
Medical Imaging Informatics. New York: Springer, 2010, pp 17-53.
[6] A. P. Dhawan and S. Dai, “Clustering and Pattern Classification”, in
Principles And Advanced Methods in Medical Imaging And Image
Analysis. Singapore: World Scientific Publishing Co. Pte. Ltd, 2008, pp
229-265.
Authorized licensed use limited to: Universitas Indonesia. Downloaded on April 28,2023 at 09:24:02 UTC from IEEE Xplore. Restrictions apply.
978-1-5090-3352-2/16/$31.00 ©2016 IEEE 16-18 November 2016, Aston Tropicana Hotel, Bandung, Indonesia
2016 International Conference on Information Management and Technology (ICIMTech)
Page 61
[7] M. S. Hosseini and M. Zekri, “Review of Medical Image Classification
using the Adaptive Neuro-Fuzzy Inference System.,” Journal of Medical
Signals & Sensors, vol. 2, no. 1, pp. 49–60, 2012.
[8] S. A. Lashari and R. Ibrahim, “A Framework for Medical Images
Classification Using Soft Set,” in Procedia Technology ICEEI 2013, pp.
548–556.
[9] M. Fauzi and B. Othman, “An Overview Of Mri Brain Classification
Using Fpga Implementation,” in IEEE Symposium on Industrial
Electronics and Applications, 2010, pp. 623–628.
[10] A. P. Dhawan, “Image Segmentation and Feature Extraction”, in
Principles And Advanced Methods in Medical Imaging And Image
Analysis. Singapore: World Scientific Publishing Co. Pte. Ltd, 2008, pp
197-228.
[11] A. A. A. Setio, F. Ciompi, G. Litjens, P. Gerke, C. Jacobs, S. J. Van
Riel, M. W. Wille, M. Naqibullah, I. S. Clara, and B. Van Ginneken,
“Pulmonary nodule detection in CT images: false positive reduction
using multi-view convolutional networks,” IEEE Transactions on
Medical Imaging, vol. 35, issue 5, 1160 - 1169, 2016.
[12] D. Mittal and A. Rani, “Detection and Classification of Focal Liver
Lesions using Support Vector Machine Classifiers,” Journal of
Biomedical Engineering and Medical Imaging, vol. 3, issue 1, pp. 21-34,
2016.
[13] G. Van Tulder and M. De Bruijne, “Combining Generative and
Discriminative Representation Learning for Lung CT Analysis with
Convolutional Restricted Boltzmann Machines,” IEEE Transactions on
Medical Imaging, vol. 35, issue 5, pp. 1262 - 1272, 2016.
[14] J. Hong, J. Vicory, J. Schulz, M. Styner, J. S. Marron, and S. M. Pizer,
“Non-Euclidean classification of medically imaged objects via s-reps
R,” Medical Image Analysis, vol. 31, pp. 37–45, 2016.
[15] K. Seetharaman and S. Sathiamoorthy, “A unified learning framework
for content based medical image retrieval using a statistical model,”
Journal of King Saud University – Computer and Information Sciences,
vol. 28, no. 1, pp. 110–124, 2016.
[16] K. Sirinukunwattana, S. E. A. Raza, Y. Tsang, D. R. J. Snead, I. A. Cree,
N. M. Rajpoot, “Locality Sensitive Deep Learning for Detection and
Classification of Nuclei in Routine Colon Cancer Histology Images,”
IEEE Transactions on Medical Imaging, vol. 35, issue 5, pp. 1–12,
2016.
[17] M. Anthimopoulos, S. Christodoulidis, L. Ebner, A. Christe, and S.
Mougiakakou, “Lung Pattern Classification for Interstitial Lung
Diseases Using a Deep Convolutional Neural Network,” IEEE
Transactions On Medical Imaging, vol. 35, no. 5, pp. 1207–1216, 2016.
[18] M. J. J. P. Van Grinsven, B. Van Ginneken, C. B. Hoyng, T. Theelen,
and I. S. Clara, “Fast convolutional neural network training using
selective data sampling: Application to hemorrhage detection in color
fundus images,” IEEE Transactions on Medical Imaging, vol. 35, issue
5, pp. 1273 - 1284, 2016.
[19] Q. Dou, S. Member, H. Chen, S. Member, L. Yu, L. Zhao, and J. Qin,
“Automatic Detection of Cerebral Microbleeds from MR Images via 3D
Convolutional Neural Networks,” IEEE Transactions on Medical
Imaging, vol. 11, no. 4, pp. 1–14, 2016.
[20] A. Masood and A. Al-jumaily, “Semi Advised SVM with Adaptive
Differential Evolution Based Feature Selection for Skin Cancer
Diagnosis,” Journal of Computer and Communications, vol. 3, pp. 184-
190, 2015.
[21] F. Khalvati, A. Wong, and M. A. Haider, “Automated prostate cancer
detection via comprehensive multi-parametric magnetic resonance
imaging texture feature models,” BMC Medical Imaging, pp. 1–14,
2015.
[22] K. Chung, E. T. Scholten, M. Oudkerk, P. A. De Jong, M. Prokop, and
B. Van Ginneken, “Automatic Classification of Pulmonary Peri-Fissural
Nodules in Computed Tomography Using an Ensemble of 2D Views
And A Convolutional Neural Network Out-of-The-Box”, Medical Image
Analysis, vol. 26, issue 1, pp. 195–202, 2015.
[23] V. Gopalakrishnan, P. G. Menon, and S. Madan, “cMRI-BED: A novel
informatics framework for cardiac MRI biomarker extraction and
discovery applied to pediatric cardiomyopathy classification,”
BioMedical Engineering, vol. 14, suppl 2, pp. 1–16, 2015.
[24] Y. Iwahori, A. Hattori, Y. Adachi, M. K. Bhuyan, and J. Robert,
“Automatic Detection of Polyp Using Hessian Filter and HOG
Features,” in Procedia Computer Science International Conference in
Knowledge Based and Intelligent Information and Engineering Systems
- KES2015, pp. 730–739, 2015.
[25] Y. Song, W. Cai, H. Huang, Y. Zhou, Y. Wang, and D. D. Feng,
“Locality-constrained Subcluster Representation Ensemble for Lung
Image Classification,” Medical Image Analysis, vol. 22, no. 1, pp. 102-
113, 2015.
[26] B. Manju, K. Meenakshy, and R. Gopikakumari, “Prostate Disease
Diagnosis from CT images using GA optimized SMRT based Texture
Features,” in Procedia Computer Science ICICT 2014, pp. 1692–1699.
[27] G. N. Balaji, T. S. Subashini, and N. Chidambaram, “Automatic
classification of Cardiac Views in Echocardiogram using Histogram and
Statistical Features,” in Procedia Computer Science ICICT 2014, pp.
1569–1576.
[28] L. Ai, X. Gao, and J. Xiong, “Application of mean-shift clustering to
Blood oxygen level dependent functional MRI activation detection,”
BMC Medical Imaging, vol. 14, no. 6, pp. 1–10, 2014.
[29] S. Yazdani, R. Yusof, A. Riazi, and A. Karimian, “Magnetic resonance
image tissue classification using an automatic method,” Diagnostic
Pathology, vol. 9, pp. 1–16, 2014.
[30] L. Tan, Y. Chen, T. C. Maloney, M. M. Caré, S. K. Holland, and L. J.
Lu, “Combined analysis of sMRI and fMRI imaging data provides
accurate disease markers for hearing impairment,” NeuroImage:
Clinical, vol. 3, pp. 416–428, 2013.
[31] R. Tomari, W. Nurshazwani, W. Zakaria, M. Mahadi, A. Jamil, F.
Mohd, N. Farhan, and N. Fuad, “Computer Aided System for Red Blood
Cell Classification in Blood Smear Image,” in Procedia Computer
Science ConfPRIDE 2013-2014, pp. 206–213.
[32] A. Andrade, J. S. Silva, J. Santos, and P. Belo-soares, “Classifier
Approaches for Liver Steatosis using Ultrasound Images,” in Procedia
Technology HCIST 2012, pp. 763–770.
[33] J. Miguel, P. Dias, C. Manta, A. Luís, and S. Cruz, “Evaluation of retinal
image gradability by image features classification,” in Procedia
Technology HCIST 2012, pp. 865–875.
[34] M. A. Dabbah, J. Graham, I. N. Petropoulos, M. Tavakoli, and R. A.
Malik, “Automatic analysis of diabetic peripheral neuropathy using
multi-scale quantitative morphology of nerve fibres in corneal confocal
microscopy imaging,” Medical Image Analysis, vol. 15, no. 5, pp. 738–
747, 2011.
[35] E. J. Ciaccio, C. A. Tennyson, G. Bhagat, S. K. Lewis, and P. H. R.
Green, “Classification of videocapsule endoscopy image patterns:
comparative analysis between patients with celiac disease and normal
individuals,” BioMedical Engineering, vol. 9, pp. 1–12, 2010.
[36] H. Wang and B. Fei, “A modified fuzzy C-means classification method
using a multiscale diffusion filtering scheme,” Medical Image Analysis,
vol. 13, no. 2, pp. 193–202, 2009.
[37] M. Mete, L. Hennings, H. J. Spencer, and U. Topaloglu, “Automatic
identification of angiogenesis in double stained images of liver tissue,”
BMC Informatics, vol. 10, suppl 11, pp. 1–14, 2009.
[38] R. Marée, P. Geurts, and L. Wehenkel, “Random subwindows and
extremely randomized trees for image classification in cell biology,”
BMC Cell Biology, vol. 8, suppl 1, pp. 1–12, 2007.
[39] M. Niemeijer, M. D. Abra, and B. Van Ginneken, “Image structure
clustering for image quality verification of color retina images in
diabetic retinopathy screening,” Medical Image Analysis, vol. 10, pp.
888–898, 2006.
[40] S. P. Awate, T. Tasdizen, N. Foster, and R. T. Whitaker, “Adaptive
Markov modeling for mutual-information-based , unsupervised MRI
brain-tissue classification,” Medical Image Analysis, vol. 10, pp. 726–
739, 2006.
[41] P. Suetens, Fundamentals of Medical Imaging, 2nd Edition, New York:
Cambridge University Press, 2009.
[42] A.P. Dhawan, Medical Image Analysis, New Jersey: John Wiley &
Sons, Inc, 2011.
Authorized licensed use limited to: Universitas Indonesia. Downloaded on April 28,2023 at 09:24:02 UTC from IEEE Xplore. Restrictions apply.
... Image scene classification is one of the most critical components of computer vision. Over the years, it has evolved significantly and found applications in various fields, including automatic surveillance systems [1], self-driving vehicles [2], medical imaging [3], and satellite image analysis [4]. The ability to classify scenes accurately is crucial for extracting meaningful information from visual data. ...
Article
Full-text available
This paper proposes a pooling-based Bag-of-Visual-Words for geographic image classification. Traditional image classification methods need a higher consumption of computational resources and slow convergence speed. In contrast, deep learning-based methods are highly demanded for large amounts of labeled data and computational resources. To address these problems, this paper introduces pooling operations, combined with the Bag-of-Visual-Words, to propose an efficient and resource-saving geographic image classification method. In the feature extraction stage, this paper combines SIFT features with maximum pooling and average pooling to generate a new feature space to capture the local and global information of the image more comprehensively. Experimental results show that the pooling-based Bag-of-Visual-Words not only improves the classification accuracy but also significantly reduces the consumption of computational resources and accelerates the convergence speed when dealing with geographic image classification. The results show that the method in this paper outperforms the original Bag-of-Visual-Words model and Spatial Pyramid Matching-based Bag-of-Visual-Words in terms of accuracy and computational efficiency, demonstrating its potential in geographic image classification with the Multi-temporal Scene Dataset of Wuhan University.
... These modalities may include magnetic resonance, © Little 2543 ionizing radiation, optical methods, nuclear medicine, ultrasound radiations, etc., functioning as one-on-one modular media for disease diagnosis and patient treatment. Each modular media has particular features and responses to the structure of the human body [35]. It generates modalities that facilitate the capture of images of the internal structure of the human body or image samples, a feat not achievable with human naked-eye visibility [36]. ...
Article
Full-text available
Medical image analysis is the most prominent research area of digital image processing. In digital image processing, medical images are more sensitive and informative. Data loss in medical image processing harms patients and the healthcare system. Therefore, several medical image analysis techniques play a significant role in ensuring lossless information during disease diagnosis. The visual clarity of an image depends upon its high resolution. Therefore, high resolution is crucial in medical image analysis. The high-resolution problem occurs due to spatial resolution limitations. These limitations can be attributed to hardware limitations, low radioactivity doses, or the acquisition time of a specific image. To avoid these limitations, several high-resolution image restoration (or reformation) techniques are available. These techniques include image compression, histogram equalization, edge detection, feature extraction, image synthesis, and noise reduction, among others. Recently, different AI-based techniques, like machine learning (ML) and deep learning (DL), have been considered booming technologies for high-resolution image reformation. Therefore, we felt it was necessary to review various AI learning-based techniques and their applications in the reformation of high-resolution digital images. The digital images may include various types, such as medical images or those from the Indian Historical Society. In this article, we first shed light on different modalities of medical image analysis techniques with their image acquisition properties. We then apply machine learning and deep learning AI approaches to the image acquisition properties to address the issue of poor visual quality in the images. This serves as evidence that ML and DL models must modify the component configuration of medical image modalities to reform high-resolution images. To improve the quality of low-resolution images, we are building new models for histogram equalization, compression, decompression, and noise reduction based on this comprehensive review.
... From the perspective of Computer Aided Diagnosis (CAD), due to the similarity of features in MRI data, Deep Learning (DL) has outperformed traditional Machine Learning (ML) in classifying neurological disorders Jo et al. (2019); Zhu et al. (2019); Rashid et al. (2020); Salehi et al. (2020); Zhang et al. (2020); Zhao and Zhao (2021); Abrol et al. (2021). With the help of DL, the structural (Mateos-Pérez et al. 2018;Bhatele and Bhadauria 2020;Falkai et al. 2022 ) and functional Rashid et al. (2020) states of the brain MRI can very well be examined and analyzed (Kumar and Karnan 2014;Liu et al. 2014;Somasundaram and Genish 2015;Miranda et al. 2016;Arora and Dhir 2017;Valliani et al. 2019;Mazurowski et al. 2019;Francis and Babu 2019;Tyagi 2019;DS 2019;Yao et al. 2020;Zhang et al. 2020;Wang et al. 2021;Zhao and Zhao 2021;Lima et al. 2022;Behrad and Abadeh 2022;Wang et al. 2022 ), assisting neurologists to reach an optimal diagnosis (Arbabshirani et al. 2017;Calhoun and Sui 2016). Substantial research has already been done toward the development of CAD systems to assist neurologists in reaching precise diagnoses (Mårtensson et al. 2020). ...
Article
Full-text available
The automatic diagnosis of neurological disorders using Magnetic Resonance Imaging (MRI) is a widely researched problem. MRI is a non-invasive and highly informative imaging modality, which is one of the most widely accepted and used neuroimaging modalities for visualizing the human brain. The advent of tremendous processing capabilities, multi-modal data, and deep-learning techniques has enabled researchers to develop intelligent, sufficiently accurate classification methods. A comprehensive literature review has revealed extensive research on the automatic diagnosis of neurological disorders. However, despite numerous studies, a systematically developed framework is lacking, that relies on a sufficiently robust dataset or ensures reliable accuracy. To date, no consolidated framework has been established to classify multiple diseases and their subtypes effectively based on various types and their planes of orientation in structural and functional MR images. This systematic review provides a detailed and comprehensive analysis of research reported from 2000 to 2023. Systems developed in prior art have been categorized according to their disease diagnosis capabilities. The datasets employed and the tools developed are also summarized to assist researchers to conduct further studies in this crucial domain. The contributions of this research include facilitating the design of a unified framework for multiple neurological disease diagnoses, resulting in the development of a generic assistive tool for hospitals and neurologists to diagnose disorders precisely and swiftly thus potentially saving lives, in addition to increasing the quality of life of patients suffering from neurodegenerative disorders.
... The imaging data generated by hospitals has significantly expanded their medical image collections through the use of diagnostic and information-gathering equipment. These images, especially X-ray, CT, endoscopy, MRI, PET, fMRI, and others, are crucial for illness investigation and accurate patient diagnosis [1][2][3]. Hence, it is vital that retrieval of such images be done in a quick and reliable manner. This task is performed by the 'image retrieval system,' which retrieves medical images from their storage repository. ...
Article
Full-text available
The rapid expansion of medical imaging repositories in hospitals has introduced significant challenges in managing and retrieving relevant data, which may contribute to diagnostic errors. Content-based medical image retrieval (CBMIR) offers a solution to these challenges by enabling efficient querying of vast datasets. This research introduces an efficient method, ResNetFuse, which leverages pre-trained deep convolutional neural networks (DCNNs), ResNet-18 and ResNet-50, for feature extraction. In ResNetFuse, the features from both networks are fused via concatenation, resulting in substantial improvements in retrieval performance. However, this fusion increases the dimensionality of the features, that leads to increase in the storage and time for retrieval process. To address the high dimensionality issue, here we used t-distributed stochastic neighbour embedding (t-SNE). The proposed ResNetFuse + t-SNE method is rigorously evaluated on the KVASIR benchmark dataset. Experimental results demonstrate that ResNetFuse + t-SNE surpasses state-of-the-art techniques across performance metrics, achieving a mean average precision (mAP) of 96.15% for the retrieval of 10 images. Additionally, the method achieves an 87.5% reduction in feature dimensionality compared to ResNetFuse alone, facilitating more compact and efficient image indexing without sacrificing retrieval accuracy. These findings underscore the efficacy of ResNetFuse + t-SNE in improving retrieval performance while reducing computational complexity, making it particularly suitable for resource-constrained environments.
... Therefore, with the recent substantial breakthroughs in this field, designing effective models for recognition and classification is imperative. Machine learning is one of the most commonly used fields of artificial intelligence [4,5,6,7]. An individual's ability to learn on their own is dependent on their past experiences [8]. ...
Article
Full-text available
This research aims to develop an image classification model by integrating long short-term memory (LSTM) with a convolutional neural network (CNN). LSTM, which is a type of neural network, can retain and retrieve long-term dependencies and improves the feature extraction capabilities of CNN when used in a multi-layer setting. The proposed approach outperforms typical CNN classifiers in image classification. The model’s high accuracy is due to the data passing through two stages and multiple layers: first the LSTM layer, followed by the CNN layer for accurate classification. Convolutional and recurrent neural networks are combined in the recommended model, which demonstrates exceptional performance on various classification tasks. The model achieved a training accuracy of 0.9899 and testing accuracy of 0.9463 using real data, which indicates its success and applicability compared with other models.
Article
Full-text available
In today’s world, cancer stands out as one of the most perilous diseases, caused by the uncontrolled proliferation of cells within the human body. Early detection is paramount to ensuring that patients receive the necessary medical intervention in a timely manner. Recently, deep learning techniques, particularly convolutional neural networks, have proven to be incredibly effective in developing computer-aided diagnosis systems due to their remarkable accuracy in analyzing medical images. However, the process of training these neural networks from scratch is often complex and requires significant computational resources. Transfer learning has emerged as a powerful solution to overcome this challenge. This study examines the fundamental concepts of machine learning and deep learning-based computer-aided diagnostic systems. It underscores the significant role of transfer learning in enhancing diagnostic accuracy. It also illustrates the various transfer learning models employed to diagnose various cancer forms, including skin cancer, brain tumors, breast cancer, lung cancer, leukemia, prostate cancer, bladder cancer, and cervical cancer. This paper summarizes 151 studies conducted in recent years. In the end, the article offers a thorough discussion of the research findings, overall conclusions, and directions for future work.
Article
Full-text available
Image classification is a fundamental task in computer science, underpinning various applications such as object detection, face recognition, and object interaction analysis. The concept holds significant value due to its wide-ranging applications across multiple fields. Traditional methods for image classification, however, have been limited by their slow processing speed, rigidity, and high costs. The integration of deep learning models, particularly Convolutional Neural Networks (CNNs) and Support Vector Machines (SVMs), has revolutionized this process, enabling the development of automated, fast, and practical systems. These advanced models are now employed in diverse areas, including biomedical science, remote sensing, and business management, thanks to their ability to achieve high accuracy across a broad spectrum of scenarios. Training these models involves the use of well-known datasets like Canadian Institute for Advanced Research (CIFAR) and Modified National Institute of Standards and Technology (MNIST), which provide the necessary data for optimization and validation. The paper examines the structure, functionality, advantages, and limitations of CNNs and SVMs in the context of image classification, demonstrating that deep learning-driven classification is now a mainstream research focus. This study highlights the transformative impact of these models and provides insights into their future potential.
Article
Full-text available
We propose a novel Computer-Aided Detection (CAD) system for pulmonary nodules using multi-view convolutional networks (ConvNets), for which discriminative features are automatically learnt from the training data. The network is fed with nodule candidates obtained by combining three candidate detectors specifically designed for solid, subsolid, and large nodules. For each candidate, a set of 2-D patches from differently oriented planes is extracted. The proposed architecture comprises multiple streams of 2-D ConvNets, for which the outputs are combined using a dedicated fusion method to get the final classification. Data augmentation and dropout are applied to avoid overfitting. On 888 scans of the publicly available LIDCIDRI dataset, our method reaches high detection sensitivities of 85.4% and 90.1% at 1 and 4 false positives per scan, respectively. An additional evaluation on independent datasets from the ANODE09 challenge and DLCST is performed. We showed that the proposed multi-view ConvNets is highly suited to be used for false positive reduction of a CAD system.
Article
Background Pediatric cardiomyopathies are a rare, yet heterogeneous group of pathologies of the myocardium that are routinely examined clinically using Cardiovascular Magnetic Resonance Imaging (cMRI). This gold standard powerful non-invasive tool yields high resolution temporal images that characterize myocardial tissue. The complexities associated with the annotation of images and extraction of markers, necessitate the development of efficient workflows to acquire, manage and transform this data into actionable knowledge for patient care to reduce mortality and morbidity. Methods We develop and test a novel informatics framework called cMRI-BED for biomarker extraction and discovery from such complex pediatric cMRI data that includes the use of a suite of tools for image processing, marker extraction and predictive modeling. We applied our workflow to obtain and analyze a dataset of 83 de-identified cases and controls containing cMRI-derived biomarkers for classifying positive versus negative findings of cardiomyopathy in children. Bayesian rule learning (BRL) methods were applied to derive understandable models in the form of propositional rules with posterior probabilities pertaining to their validity. Popular machine learning methods in the WEKA data mining toolkit were applied using default parameters to assess cross-validation performance of this dataset using accuracy and percentage area under ROC curve (AUC) measures. Results The best 10-fold cross validation predictive performance obtained on this cMRI-derived biomarker dataset was 80.72% accuracy and 79.6% AUC by a BRL decision tree model, which is promising from this type of rare data. Moreover, we were able to verify that mycocardial delayed enhancement (MDE) status, which is known to be an important qualitative factor in the classification of cardiomyopathies, is picked up by our rule models as an important variable for prediction. Conclusions Preliminary results show the feasibility of our framework for processing such data while also yielding actionable predictive classification rules that can augment knowledge conveyed in cardiac radiology outcome reports. Interactions between MDE status and other cMRI parameters that are depicted in our rules warrant further investigation and validation. Predictive rules learned from cMRI data to classify positive and negative findings of cardiomyopathy can enhance scientific understanding of the underlying interactions among imaging-derived parameters.
Article
In the present work, two computer aided diagnostic systems are designed to detect and classify focal liver lesions such as Cyst, Hemangioma, Hepatocellular carcinoma and Metastases. The work evaluates clinically acquired ultrasound image database. Database contains 111 liver images comprising 95 images of focal lesions and 16 images of normal liver. Images are enhanced and manually segmented into 800 non-overlapping segmented-regions-of-interest. Afterwards, 208 textual features are extracted from each segmented-regions-of-interest. First diagnostic system is designed with one-against-one multiclass support vector machine classification approach showing 93.1% (512/550) overall accuracy on test dataset. Second system is designed with tree structured approach using four binary support vector machine classifiers showing 86.9% (478/550) overall accuracy on test dataset. Out of these two, one-against-one approach based diagnostic system outperforms the neural network based diagnostic system designed for the same purpose by providing 96.6% classification accuracy for typical cases and 85.3% for atypical cases.
Chapter
Clustering is a method to arrange data points into groups or clusters based on a predefined similarity criterion. Classification maps the data points or their representative features into predefined classes to help the interpretation of the input data. There are several methods available for clustering and classification for computeraided diagnostic or decision making systems for medical applications. This chapter reviews some of the clustering and classification methods using deterministic as well as fuzzy approaches for data analysis. © 2008 by World Scientific Publishing Co. Pte. Ltd. All rights reserved.
Chapter
Medical image segmentation tasks are important to visualize features of interests such as lesions with boundary and volume information. Similar information is required in the computerized quantitative analysis and classification for diagnostic evaluation and characterization. This chapter presents some of the most effective and commonly used edge and region segmentation methods. Statistical quantitative features from gray level distribution, segmented regions, and texture in the image are also presented. © 2008 by World Scientific Publishing Co. Pte. Ltd. All rights reserved.
Article
Automated tissue characterization is one of the most crucial components of a computer aided diagnosis (CAD) system for interstitial lung diseases (ILDs). Although much research has been conducted in this field, the problem remains challenging. Deep learning techniques have recently achieved impressive results in a variety of computer vision problems, raising expectations that they might be applied in other domains, such as medical image analysis. In this paper, we propose and evaluate a convolutional neural network (CNN), designed for the classification of ILD patterns. The proposed network consists of 5 convolutional layers with 2×2 kernels and LeakyReLU activations, followed by average pooling with size equal to the size of the final feature maps and three dense layers. The last dense layer has 7 outputs, equivalent to the classes considered: healthy, ground glass opacity (GGO), micronodules, consolidation, reticulation, honeycombing and a combination of GGO/reticulation. To train and evaluate the CNN, we used a dataset of 14696 image patches, derived by 120 CT scans from different scanners and hospitals. To the best of our knowledge, this is the first deep CNN designed for the specific problem. A comparative analysis proved the effectiveness of the proposed CNN against previous methods in a challenging dataset. The classification performance (~85.5%) demonstrated the potential of CNNs in analyzing lung patterns. Future work includes, extending the CNN to three-dimensional data provided by CT volume scans and integrating the proposed method into a CAD system that aims to provide differential diagnosis for ILDs as a supportive tool for radiologists.
Article
Classifying medically imaged objects, e.g., into diseased and normal classes, has been one of the important goals in medical imaging. We propose a novel classification scheme that uses a skeletal representation to provide rich non-Euclidean geometric object properties. Our statistical method combines distance weighted discrimination (DWD) with a carefully chosen Euclideanization which takes full advantage of the geometry of the manifold on which these non-Euclidean geometric object properties (GOPs) live. Our method is evaluated via the task of classifying 3D hippocampi between schizophrenics and healthy controls. We address three central questions. 1) Does adding shape features increase discriminative power over the more standard classification based only on global volume? 2) If so, does our skeletal representation provide greater discriminative power than a conventional boundary point distribution model (PDM)? 3) Especially, is Euclideanization of non-Euclidean shape properties important in achieving high discriminative power? Measuring the capability of a method in terms of area under the receiver operator characteristic (ROC) curve, we show that our proposed method achieves strongly better classification than both the classification method based on global volume alone and the s-rep-based classification method without proper Euclideanization of non-Euclidean GOPs. We show classification using Euclideanized s-reps is also superior to classification using PDMs, whether the PDMs are first Euclideanized or not. We also show improved performance with Euclideanized boundary PDMs over non-linear boundary PDMs. This demonstrates the benefit that proper Euclideanization of non-Euclidean GOPs brings not only to s-rep-based classification but also to PDM-based classification.
Article
The choice of features greatly influences the performance of a tissue classification system. Despite this, many systems are built with standard, predefined filter banks that are not optimized for that particular application. Representation learning methods such as restricted Boltzmann machines may outperform these standard filter banks because they learn a feature description directly from the training data. Like many other representation learning methods, restricted Boltzmann machines are unsupervised and are trained with a generative learning objective; this allows them to learn representations from unlabeled data, but does not necessarily produce features that are optimal for classification. In this paper we propose the convolutional classification restricted Boltzmann machine, which combines a generative and a discriminative learning objective. This allows it to learn filters that are good both for describing the training data and for classification. We present experiments with feature learning for lung texture classification and airway detection in CT images. In both applications, a combination of learning objectives outperformed purely discriminative or generative learning, increasing, for instance, the lung tissue classification accuracy by 1 to 8 percentage points. This shows that discriminative learning can help an otherwise unsupervised feature learner to learn filters that are optimized for classification.
Article
Convolutional neural networks (CNNs) are deep learning network architectures that have pushed forward the state-of-the-art in a range of computer vision applications and are increasingly popular in medical image analysis. However, training of CNNs is time-consuming and challenging. In medical image analysis tasks, the majority of training examples are easy to classify and therefore contribute little to the CNN learning process. In this paper, we propose a method to improve and speed-up the CNN training for medical image analysis tasks by dynamically selecting misclassified negative samples during training. Training samples are heuristically sampled based on classification by the current status of the CNN. Weights are assigned to the training samples and informative samples are more likely to be included in the next CNN training iteration. We evaluated and compared our proposed method by training a CNN with (SeS) and without (NSeS) the selective sampling method. We focus on the detection of hemorrhages in color fundus images. A decreased training time from 170 epochs to 60 epochs with an increased performance - on par with two human experts - was achieved with areas under the receiver operating characteristics curve of 0.894 and 0.972 on two data sets. The SeS CNN statistically outperformed the NSeS CNN on an independent test set.
Article
Detection and classification of cell nuclei in histopathology images of cancerous tissue stained with the standard hematoxylin and eosin stain is a challenging task due to cellular heterogeneity. Deep learning approaches have been shown to produce encouraging results on histopathology images in various studies. In this paper, we propose a Spatially Constrained Convolutional Neural Network (SC-CNN) to perform nucleus detection. SC-CNN regresses the likelihood of a pixel being the center of a nucleus, where high probability values are spatially constrained to locate in the vicinity of the center of nuclei. For classification of nuclei, we propose a novel Neighboring Ensemble Predictor (NEP) coupled with CNN to more accurately predict the class label of detected cell nuclei. The proposed approaches for detection and classification do not require segmentation of nuclei. We have evaluated them on a large dataset of colorectal adenocarcinoma images, consisting of more than 20,000 annotated nuclei belonging to four different classes. Our results show that the joint detection and classification of the proposed SC-CNN and NEP produces the highest average F1 score as compared to other recently published approaches. Prospectively, the proposed methods could offer benefit to pathology practice in terms of quantitative analysis of tissue constituents in whole-slide images, and could potentially lead to a better understanding of cancer.