2019 International Conference on Computing, Mathematics and Engineering Technologies – iCoMET 2019

Medical Imaging using Machine Learning and

Deep Learning Algorithms: A Review
*
Jahanzaib Latif, Chuangbai Xiao, *Azhar Imran, Shanshan Tu
Faculty of Information Technology
Beijing University of Technology
Beijing, China
Email: [email protected], [email protected]

Abstract—Machine and deep learning algorithms are rapidly position of medical imaging. The digital image processing
growing in dynamic research of medical imaging. Currently, offers significant effect on decision-making procedure based
substantial efforts are developed for the enrichment of medical on some predictions. It gives better features extraction and
imaging applications using these algorithms to diagnose the accuracy. The procedure of functioning assessment is
errors in disease diagnostic systems which may result in complicated and contains numerous diverse properties [5] [6].
extremely ambiguous medical treatments. Machine and deep The digital image processing techniques are implanted in many
learning algorithms are important ways in medical imaging to different computer systems. The authentication of image
predict the symptoms of early disease. Deep learning techniques, processing approaches is essential that gives an implementation
in specific convolutional networks, have promptly developed a
of specific procedures which provides influence on the
methodology of special for investigating medical images. It uses
performance of these systems. Therefore, it brings decisions
the supervised or unsupervised algorithms using some specific
standard dataset to indicate the predictions. We survey image and actions based on approaches in medical imaging. It
classification, object detection, pattern recognition, reasoning etc. delivers a many rudimentary and refined image analysis and
concepts in medical imaging. These are used to improve the visualization tools [7]. The artificial intelligence is the main
accuracy by extracting the meaningful patterns for the specific domain and machine learning and deep learning works under
disease in medical imaging. These ways also indorse the decision- this domain as shown in Fig. 1. The AI is the major field to
making procedure. The major aim of this survey is to highlight display human intelligence in a machine, machine learning is
the machine learning and deep learning techniques used in used to achieve artificial intelligence, while deep learning is a
medical images. We intended to provide an outline for technique used to implement machine learning [8].
researchers to know the existing techniques carried out for
medical imaging, highlight the advantages and drawbacks of The various steps are performed on medical images before
these algorithms, and to discuss the future directions. For the the detection of output. Initially, the medical image is inserted
study of multi-dimensional medical data, machine and deep as input to the machine and deep learning algorithms. After
learning provide a commendable technique for creation of that, the image is divided into different segments to zoom the
classification and automatic decision making. This paper interested area. Then, the features are extracted from these
provides a survey of medical imaging in the machine and deep segments through information retrieval techniques. Next. The
learning methods to analyze distinctive diseases. It carries desired features are selected and the noise is removed.
consideration concerning the suite of these algorithms which can
be used for the investigation of diseases and automatic decision-
making.

Keywords—Medical imaging; Machine learning; Deep

learning; Image enhancement; Information retrieval;

I. INTRODUCTION
Machine and deep learning algorithms play an important
role to train the computer system as an expert which can be
used further for prediction and decision making. Machine
learning is the field of study that provides computers the ability
to learn without being explicitly programmed [1]. Deep
learning is a type of machine learning that empowers systems
to gain for a fact and comprehend the world regarding a
pecking order of ideas [2]. These fields bring intelligence into a
computer that can extract the patterns according to the specific
data and then process for automatic reasoning [3] [4]. Medical
imaging is the rapidly growing research area that is used to
Fig. 1. Artificial Intelligence, Machine Learning and Deep Learning domain
diagnose a disease for early treatment. The function of image
processing in the health domain is relative to the growing

978-1-5386-9509-8/19/$31.00 ©2019 IEEE

Authorized licensed use limited to: REVA UNIVERSITY. Downloaded on June 19,2025 at 10:23:28 UTC from IEEE Xplore. Restrictions apply.
 2019 International Conference on Computing, Mathematics and Engineering Technologies – iCoMET 2019

Finally, the classifier is used to classify the extracted data optimality of substances to gain tags for the goal. Such as
and make predictions based on this classification. These steps budgets functions in an organization.
are used in every experiment of machine learning. The
supervised, semi-supervised, unsupervised, reinforcement and E. Reinforcement learning
active learning algorithms are the main categories of machine In Reinforcement learning the trained data is provided only
learning. Further, the deep learning methods are basically as a response to the program's activities in a self-motivated
advanced phase of machine learning algorithms that classify situation, such as to drive a vehicle or playing a video game.
data and predict more accurately using neural network [9] [10].
F. Evolutionary Learning
A. Supervised learning
It is mainly used in the biological field to learn biological
It gives a training set of instances with appropriate organisms and to predict their survival rate and the casual of
objectives to a computer system. Taking this training set the offsprings. We can use this model by using the knowledge
system give response accurately on given possible inputs. The of fitness, to predict how to correct the result.
classification and regression are the categories of Supervised
Learning.
G. Deep learning
• The inputs are distributed into different classes using This is the advance phase of machine learning which
classification methods, and the trained system must mainly uses neural networks for learning and prediction of
generate actions that allocate hidden inputs to these data. It is a group of different algorithms. These are used to
classes. This is called multi labeling process. The spam design complex generalize system that can take any type of
purifying is the case of classification, in which the problems and give predictions. It uses the deep graph with
emails are classified into "spam" and, "not spam". numerous processing layer, made up of many linear and
• The regression is a supervised technique in which the nonlinear conversion [11].
outcomes are continuous rather than discrete. The Nowadays, in medical disciplines, illness diagnostic
regression predictions are evaluated using root mean assessment is a severe mission. It is the key task to comprehend
squared error (RMSE), unlike classification predictions the accurate diagnosis of patients by medical inspection and
in which accuracy is used as a performance measure. evaluation. The healthcare domain produces a large amount of
data about medical evaluation, a statement concerning the
B. Unsupervised learning patient, treatment, supplements, prescription etc. The main
The system will take the decision by itself rather train on issue is that the quality of these reports gives an effect of
the basis of some dataset. No labeling is given to the system association due to unsuitable managing of the data [12]. The
that can be used for predictions. Unsupervised learning can be improvements of this data are essential to mine and process
used to retrieve the hidden pattern with the help of feature these medical reports elegantly and efficiently. There are
learning of the given data. different types of machine learning algorithms are available
that are used to use specific classifier for distributing data on
• The clustering is an unsupervised learning approach the basis of their characteristics. The dataset can be separated
that is used to divide the inputs into clusters. These into two or more than two classes. These types of classifiers are
clusters are not identified earlier. It builds groups on embedded for medical data examination and disease discovery.
the basis of resemblance. Firstly, machine learning algorithms were proposed and used to
notice medical data sets. Today, for well-organized
C. Semi-supervised learning investigation of medical data the machine learning provides
In Semi-supervised learning, the system is assumed to be numerous techniques. The systems for data gathering and
partial training data. This type of training is used with some inspection are located in almost all new hospitals that are used
trained data that can target some missing results. This type for the collection and sharing of data. It is used for correcting
algorithm is used on untagged data for training commitment. diagnostic of different diseases by medical imaging. To use an
The semi-supervised learning algorithm trained on both labeled algorithm, the accurate diagnostic patient documentation is
and unlabeled data and this learning exhibits the features the inserted in a system and outcomes can be spontaneously
features of both the unsupervised-learning and supervised acquired from the preceding cracked cases. The pattern
learning algorithms. recognition is the idea of machine learning that extracts
features from medical images about any diseases to predict and
D. Active learning generate conclusions for diagnosis and to plan treatment [13]
[14]. There are different steps of machine learning and deep
In Active learning, the system gets that training tags only learning algorithms which are used in the medical imaging
for a restricted set of occurrences. It is used to enhance its domain as shown in Fig. 2.

978-1-5386-1370-2/18/$31.00 ©2019 IEEE

Authorized licensed use limited to: REVA UNIVERSITY. Downloaded on June 19,2025 at 10:23:28 UTC from IEEE Xplore. Restrictions apply.
 2019 International Conference on Computing, Mathematics and Engineering Technologies – iCoMET 2019

analysis, support vector machines, decision trees, etc. In [17],

the author used machine learning approaches in medical image
evaluation. In specific, they used local binary patterns
extensively contemplated among texture descriptors. Further, a
study on new trials using several low binary patterns
descriptors of biomedical images. The dataset of neonatal facial
images for categorizing pain conditions beginning from facial
descriptions. Especially, the outcomes on the extensively
premeditated 2D-HeLa dataset and the suggested descriptor
gains the maximum implementation including all the numerous
texture descriptors. A linear support vector machine classifier
is applied on the 2D-HeLa dataset and in the PAP dataset. The
92.4 % accuracy got which is the highest values among all
other descriptors on the mentioned dataset. The neural network
technique is used in medical images to investigate the disease
details [18]. The neural network groups are retained for cancer
discovery. It is used to critic where a cell is normal with
excessive assurance where each distinct network has only two
outcomes either it will be a normal cell or cancer cell. The
predictions of these cells’ network are merged by a
predominant method, i.e. plurality voting. The results showed
that the neural network collectively accomplished a high rate of
Fig. 2. Machine and Deep Learning algorithms workflow in medical image accuracy and a low value of false negative analysis.

II. MACHINE LEARNING IN MEDICAL IMAGING The machine learning expert systems provide contrivances
for the production of premises from patients’ information.
Machine learning algorithms are very effective in medical Different rules are mined from the information of specialists to
imaging to study specific diseases. Different types of entities paradigm an expert system. The group of clinical problems that
such as lesions and organs in medical image processing can be can be used as examples, knowledge in intelligent systems may
too complicated and cannot be shown correctly by a simple achieve by machine learning approaches that can be used to
mathematical solution. In [15], the author used the pixel-based generate a methodical description of clinical characters that
investigation to analyze medical images for diseases. The pixel distinctively describe the clinical circumstances. Therefore,
analysis in machine learning appeared in medical image information can be articulated in the arrangement of simple
processing, which uses certain values in images straight away rules, or often as a decision tree. A typical example of this
instead of features extraction from chunks as input data. The category of the scheme is KARDIO, which is grown to
enactment of this method may possibly be better than that of translate ECGs [19]. In medical image analysis, the good
simple feature-based classifiers for specific problems. The standard for evaluating image feature is a statistical analysis.
image with low contrast is a challenging problem in order to The channelized Hoteling observer (CHO), is generally used
investigate its properties. The feature calculation and for specifically in nuclear medicine imaging. The channels are
segmentation is not required for pixel-based machine learning, enthused by the idea of amenable subjects in the human visual
unlike ordinary classifiers which avoid errors generated from structure. This method is used to detect image quality
inaccurate segmentation and feature calculation. The pixel evaluation and further, the CHO has defensibly and positive
analysis utilizes long training time because of the high influence on the medical imaging. The subsequent algorithm is
dimensionality of data (a large number of pixel in an image) in named a channelized SVM (CSVM). There are two medical
[16], the author targeted the low contrast medical images for physicists assessed the flaw discernibility in 100 noisy images
analysis. The furthermost efficient technique used for contrast and then the score confidence of a lesion actuality
improvement is Histogram Equalization (HE). The authors contemporary on a six-point scale. After that, a training session
proposed a technique named “Modified Histogram-Based is used to involve an extra 60 images. The human spectators
Contrast Enhancement using Homomorphic Filtering" (MH- achieved this assignment for six diverse selections of the
FIL). It used two phases handling process, in the first phase flattening filter with two dissimilar choices of the number of
global contrast is improved using histogram modification. repetitions in the OS-EM rebuilding algorithm [20].
Further, second phase homomorphic filtering is projected for
image sharpening. The low contrast chest X-ray 10 medical
images are investigated in the experiment. The MH-FIL has III. DEEP LEARNING IN MEDICAL IMAGING
minimum values in all 10 images computer to other techniques. To guide computers to learn features that can characterize
The medical image clarification is the highest responsibility of the data for the given issue. This idea lies at the foundation of
radiologists, with the assignments involving equally images several deep learning procedures. The models that comprised
with better quality and its analysis. The computer-aided design of various layers that transmute input images to give outputs
CAD has developed for several years. about the specific diseases because of cramming gradually
There are numerous machine learning methods analyzed high-level features. The better type of these models for image
through medical images, for example, linear discriminant analysis is convolutional neural networks (CNNs). The CNNs

978-1-5386-1370-2/18/$31.00 ©2019 IEEE

Authorized licensed use limited to: REVA UNIVERSITY. Downloaded on June 19,2025 at 10:23:28 UTC from IEEE Xplore. Restrictions apply.
 2019 International Conference on Computing, Mathematics and Engineering Technologies – iCoMET 2019

comprise several layers that convert the input with convolution medical images in order to provide comparison or integration
filters. The task of employing deep learning methods to the of the data obtained from a different viewpoint, time, depth,
medical field frequently use in familiarizing current and sensors etc. This is the iterative process in which we select
architectures in distinctive input formats such as three- a specific type of parameters as a standard. It is used to
dimensional data. Previously, the purposes of CNNs to big calculate the similarity parameters of two images using deep
data, full 3D convolutions and the subsequent huge number of learning algorithms [28]. The registration is used in medicine
constraints are avoided by separating the volume of Interest i.e. Computer Tomography (CT) and NMR data. This is quite
into portions [21]. helpful to attain patient information, observing tumor growth,
cure confirmation, and the comparison of the patient’s
A. Classification
information with anatomical atlases. The mutual information
1) Image classification obtained in [29] using Powell’s and Brent’s method to register
The classification if medical image is the main task in deep MR, CT is different from [30] which is used on breast MR
learning in order to investigate for clinical-related issues for images.
early treatment of the patient. The classification might be
classically or multiple images as input with a single diagnostic IV. CONCLUSION
mutable as result (disease yes or no). In these cases, each The machine learning skills are grown in past few years.
diagnostic test is a model and dataset sizes are Currently, machine learning methods are tremendously
characteristically minor related to those in computer vision. In vigorous to practical circumstances, and the structures really
[22], the fine modification evidently outdid feature extraction, advantage of the learning process. It previously pertains in the
attaining 57.6% accuracy in multiclass score evaluation of rehearsal of medical imaging, and it will perhaps cultivate at a
knee osteoarthritis against 53.4%. But, the [23] shown that by quick stride in the coming future. The use of machine learning
CNN feature retrieval performed fine-tuning in cytopathology in medical imaging has important inferences for the
image classification accuracy with 70.5% versus 69.1%. medication. It is very significant that this research area ensures
2) Object classification better care to patients. The possessions of machine learning
The object classification targets on the small interested tackles are serious to confirming that they are applied in the
chunks of the medical image. These chunks can be projected greatest real way. In the medical image analysis, the deep
into two or more classes. For better accuracy, the local learning algorithms help to categorize, classify, and enumerate
information of these chunks and global conceptual information disease patterns from image processing. It also permits to
are very important. In [24], the author used three CNNs extend analytical goals and generates prediction models of
methods of deep learning to patch the image at a different treatment for patients. The medical imaging researchers are
scale of objects. The results of these three techniques finally considering these challenges, deep learning in health care
reflected the features matrix of the overall image properties. research domain and imaging is enduring to flourish. It is
B. Detection: Organ and region improving rapidly, as deep learning is in numerous other
applications other than of health care.
The object detection and localization is the next phase after
classification. It is an important step in segmentation where REFERENCES
we can extract the importance of each object and focus only
[1] Valiant, L., G. A theory of the learnable. Commun. ACM,
the interested object and discard the noise. To challenge this
27(11):1134–1142, November 1984.
issue, a 3D data parsing approach is used using deep learning
[2] Goodfellow, I., Bengio, Y., Courville, A., & Bengio, Y.
algorithms. The author used three independent groups 2D and
(2016). Deep learning (Vol. 1). Cambridge: MIT Press.
3D MRI chunks in medical image. It is used to locate the [3] Robert, C., Machine learning, a probabilistic perspective.
regions of different related objects which focus on some 2014, Taylor & Francis.
specific diseases such heart, aortic arch, descending aorta [25,
[4] Aiken, A., Moss: A system for detecting software
26].
plagiarism.
C. Segmentation http://www/.cs.berkeley.edu/~aiken/moss.html, 2004.
The segmentation process is used to process the organs and [5] Doi, K., Computer-aided diagnosis in medical imaging:
substructures of the medical images. It is used for quantitative historical review, current status and future potential.
analysis of the clinical features. For example, cardiac or brain Computerized medical imaging and graphics, 2007. 31(4-
examination. It is also used in CAD for functions. It is the 5): p. 198-211.
identifying of specific pixels that make up the object of [6] Mahesh, M., Fundamentals of medical imaging. Medical
interest. The U-net is the combination of upsampling and Physics, 2011. 38(3): p. 1735-1735.
downsampling layers architectures. It merged the connections [7] Jannin, P., C. Grova, and C.R. Maurer, Model for
of convolution and de-convolution samples of layers [27]. defining and reporting reference-based validation
protocols in medical image processing. International
D. Registration Journal of Computer Assisted Radiology and Surgery,
The registration is the process of transforming different sets 2006. 1(2): p. 63-73.
of data into one coordinate system. It is a necessary step in

978-1-5386-1370-2/18/$31.00 ©2019 IEEE

Authorized licensed use limited to: REVA UNIVERSITY. Downloaded on June 19,2025 at 10:23:28 UTC from IEEE Xplore. Restrictions apply.
 2019 International Conference on Computing, Mathematics and Engineering Technologies – iCoMET 2019

[8] Michalski, R. S., Carbonell, J. G., & Mitchell, T. M. network. In International conference on medical image
(Eds.). (2013). Machine learning: An artificial computing and computer-assisted intervention. 2013.
intelligence approach. Springer Science & Business Springer.
Media. [22] Antony, J., et al. Quantifying radiographic knee
[9] Norris, D.J., Machine Learning: Deep Learning, in osteoarthritis severity using deep convolutional neural
Beginning Artificial Intelligence with the Raspberry Pi. networks. In Pattern Recognition (ICPR), 2016 23rd
2017, Springer. p. 211-247. International Conference on. 2016. IEEE.
[10] Jankowski, N. and M. Grochowski. Comparison of [23] Kim, E., M. Corte-Real, and Z. Baloch. A deep semantic
instances seletion algorithms i. algorithms survey. In mobile application for thyroid cytopathology. In Medical
International conference on artificial intelligence and soft Imaging 2016: PACS and Imaging Informatics: Next
computing. 2004. Springer. Generation and Innovations. 2016. International Society
[11] Schmidhuber, J., Deep learning in neural networks: An for Optics and Photonics.
overview. Neural networks, 2015. 61: p. 85-117. [24] Shen, W., et al. Multi-scale convolutional neural networks
[12] Warwick, W., et al., A framework to assess healthcare for lung nodule classification. In International Conference
data quality. The European Journal of Social & on Information Processing in Medical Imaging. 2015.
Behavioural Sciences, 2015. 13(2): p. 1730. Springer.
[13] Ghassemi, M., et al., Opportunities in Machine Learning [25] Wang, C.-W., et al., Evaluation and comparison of
for Healthcare. arXiv preprint arXiv:1806.00388, 2018. anatomical landmark detection methods for cephalometric
[14] Dua, S., U.R. Acharya, and P. Dua, Machine learning in x-ray images: a grand challenge. IEEE transactions on
healthcare informatics. 2014. medical imaging, 2015. 34(9): p. 1890-1900.
[15] Suzuki, K., Pixel-based machine learning in medical [26] De Vos, B.D., et al. 2D image classification for 3D
imaging. Journal of Biomedical Imaging, 2012. 2012: p. anatomy localization: employing deep convolutional
1. neural networks. In Medical Imaging 2016: Image
[16] Agarwal, T.K., M. Tiwari, and S.S. Lamba. Modified Processing. 2016. International Society for Optics and
histogram based contrast enhancement using Photonics.
homomorphic filtering for medical images. In Advance [27] Çiçek, Ö. et al. 3D U-Net: learning dense volumetric
Computing Conference (IACC), 2014 IEEE International. segmentation from sparse annotation. In International
2014. IEEE. Conference on Medical Image Computing and Computer-
[17] Nanni, L., A. Lumini, and S. Brahnam, Local binary Assisted Intervention. 2016. Springer.
patterns variants as texture descriptors for medical image [28] Wu, G., et al. Unsupervised deep feature learning for
analysis. Artificial intelligence in medicine, 2010. 49(2): deformable registration of MR brain images. In
p. 117-125. International Conference on Medical Image Computing
[18] Shi, Z. and L. He. Application of neural networks in and Computer-Assisted Intervention. 2013. Springer.
medical image processing. In Proceedings of the second [29] F. Maes, A. Collignon, D. Vandermeulen, G. Marchal, P.
international symposium on networking and network Suetens, Multimodality image registration by
security. 2010. Citeseer. maximization of mutual information, IEEE Transactions
[19] Bratko, I., I. Mozetič, and N. Lavrač, KARDIO: a study on Medical Imaging 16 (1997) 187–198.
in deep and qualitative knowledge for expert systems. [30] D. Rueckert, C. Hayes, C. Studholme, P. Summers, M.
1990: MIT Press. Leach, D.J. Hawkes, Non-rigid registration of breast MR
[20] Narasimhamurthy, A., An overview of machine learning images using mutual information, Proceedings of the
in medical image analysis: Trends in health informatics, Medical Image Computing and Computer-Assisted
in Classification and Clustering in Biomedical Signal Intervention MICCAI’98, Cambridge, Massachusetts,
Processing. 2016, IGI Global. p. 23-45. 1998, pp. 1144–1152
[21] Prasoon, A., et al. Deep feature learning for knee cartilage
segmentation using a triplanar convolutional neural

978-1-5386-1370-2/18/$31.00 ©2019 IEEE

Authorized licensed use limited to: REVA UNIVERSITY. Downloaded on June 19,2025 at 10:23:28 UTC from IEEE Xplore. Restrictions apply.