Volume-3,Issue-5,May2025 International Journal of Modern Science and Research Technology

ISSN NO-2584-2706

A Review on Deep Learning for Medical

Image Processing
Sandhya Dahake; Manvi Godbole
Vaishnavi Zore; ArvekshaChambhare
Department of Master in Computer Application,
G H Raisoni College of Engineering and Management,
Maharashtra, India

Abstract: Machine learning algorithms can be deeply

Deep learning is rapidly emerging, leading to embedded throughout all fields of medicine,
dramatic advances in some medical applications. ranging from drug development to clinical
The application of the Deep Learning Approach decision-making, significantly changing how
to medical image processing is quickly growing medical practice is carried out. [2] The
as a field of study to determine whether diseases quantity of images for a specific patient case
are present or not. Most Deep Learning rises significantly from a handful of two-
Applications implementations focus on digital dimensional (2D) images to hundreds with
histopathological images, computed tomography, 3D imaging and thousands with 4D dynamic
X-ray images, and mammograms. The latest imaging.
technique in machine learning is believed to be The Recent success of machine learning
inspired by deep learning algorithms and algorithms in computer vision tasks aligns
applications. This paper presents a perfectly with the growing reliance on
comprehensive review on classification, medical records [3]. The adoption of
detection, and segmentation of medical images electronic health records (EHR) increased
via Deep Learning Applications. fourfold from 11.8% to 39.6% among office-
based physicians in the US between 2007 and
Keywords: Medical image processing, Deep 2012 [4]. An incorrect or late diagnosis
Learning, Unsupervised learning, Segmentation, harms the patient. Consequently, it is optimal
Classification for medical image analysis to be performed
by an automated, precise, and effective
1. Introduction: machine learning algorithm because delayed
Recent progress in machine learning and or incorrect diagnoses can have severe
computational methods has underscored the consequences, effective and automated
importance and need for developing machine learning algorithms for medical
computerized tools to support radiologists in image analysis are becoming ever more
image analysis and diagnosis, establishing it as a valuable. Human professionals such as
crucial area of research and development in radiologists and physicians have long been
medical imaging. [1] The increasing workload interpreting these images [5].
prevents radiologists and physicians from [6] Artificial intelligence (AI) will be applied
maintaining workflow efficiency while utilizing in preventive medicine, diagnostic support,
all accessible imaging data to improve accuracy personalized medicine, innovation in therapy,
and patient care. growth of healthy life expectancy, and
nursing care services, relieving the workload
of caregivers. Medical areas where AI will
apply for real-world use shortly are genomic
medicine, diagnostic imaging assistance

IJMSRT25MAY080 www.ijmsrt.com 352

DOI: https://doi.org/10.5281/zenodo.15522486
Volume-3,Issue-5,May2025 International Journal of Modern Science and Research Technology
ISSN NO-2584-2706

(medical image analysis), diagnosis and imaging due to the high cost and time required
treatment support, and drug discovery. for expert annotations [11].
Healthcare is a large and analog business
industry and therefore one of the world's top 2) Self-supervised Learning: It refers to ML
IT firms [7]. techniques where models learn meaningful
representations from unlabelled medical
2. Methods: images, extracting valuable information
The followings are the 14 sorts of learning directly from the data without requiring
that we should be acquainted with as an AI explicit human annotations, allowing for
specialist. robust analysis even when labeled data is
scarce [12].
2.1. Learning problems:
1) Supervised Learning: This algorithm is 3) Multi-instance Learning: It is a technique
learned using vast sets of labeled medical used to classify medical images where the
images to recognize patterns and make label is associated with a "bag" of smaller
predictions regarding the presence or image patches (instances), rather than a
nature of the disease in those images, single image, allowing for the analysis of
thereby enabling computers to help with complex medical images even when only a
the diagnosis by identifying abnormal small portion of the image contains the
conditions based on earlier labeled relevant pathology, making it particularly
examples [8]. useful in scenarios where precise
localization of the target area is not readily
2) Unsupervised Learning: It applies ML available [13].
algorithms to learn patterns & structure from
unlabelled medical images so that scientists 2.3. Statistical inference:
can extract new insights from the data 1) Inductive Learning: It refers to a machine
without necessarily having manually learning approach where a model learns
annotated labels, which are often expensive general rules and patterns from a set of
and time-consuming to obtain useful for labeled medical images, allowing it to
anomaly detection finding new subtypes of make predictions on new, unseen images
diseases in big medical image collections [9]. by identifying similar characteristics,
essentially "generalizing" from the training
data to diagnose diseases or analyze
3) Reinforcement learning (RL): It is an anatomical structures based on the patterns
approach to ML in which a computer observed in the training set [14].
program, or algorithm, known as an "agent,"
learns how to maximize outcomes in a 2) Deductive Inference: It is a top-down
medical image by responding to feedback approach to analysis, where a computer
through rewards, to improve performance at system uses known information to make
complicated operations such as the detection, deductions about what is present in the
segmentation, or tracking of lesions[10]. image, much like a doctor might use a
patient's symptoms to reach a diagnosis. It
2.2. Hybrid learning problems: refers to a process where a computer
1) Semi-supervised Learning: It refers to a system uses established medical
technique where a model is trained using a knowledge and patterns in an image to
combination of both labeled & unlabelled draw specific conclusions about a patient's
medical images, allowing it to learn from a condition [15].
larger dataset even when limited labeled data is
available, which is often the case in medical 2.3.Learning techniques:

IJMSRT25MAY080 www.ijmsrt.com 353

DOI: https://doi.org/10.5281/zenodo.15522486
Volume-3,Issue-5,May2025 International Journal of Modern Science and Research Technology
ISSN NO-2584-2706

1) Transductive Learning: It is a technique strengths of different models, ensemble

where a model leverages both labeled and learning can produce more reliable
unlabelled medical images within a diagnoses and predictions from medical
specific dataset to make predictions about images like X-rays, MRIs, and CT scans
the unlabelled data, essentially "learning [20].
from the context" of the unlabelled images
to improve classification accuracy [16]. 3. Overview:

2) Multi-task Learning: It is a fundamental

learning paradigm for machine learning,
which aims to simultaneously solve multiple
related tasks to improve the performance of
individual tasks by sharing knowledge.

3) Active Learning: It is a technique where a

model strategically selects the most
informative unlabelled medical images to be
Fig.1. Overview of deep learning techniques,
annotated by experts, allowing it to learn
applications, and framework [21]
effectively with a minimal amount of labeled
data, significantly reducing the cost and time
required for training accurate diagnostic 3.1 Deep Learning Techniques:
1) Deep Neural Network (DNN): In this
models, particularly when dealing with
architecture, at least two layers are there
limited annotated medical image datasets [17].
that allow nonlinear complexities.
Classification and regression can be carried
4) Online Learning: It refers to the process of
out here. The advantage of this model is
acquiring knowledge and skills in analyzing
generally used because of its great accuracy.
medical images (like X-rays, CT scans and
The drawback is that the method of training
MRIs) through digital platforms, primarily
will not be easy since the error is
utilizing techniques like deep learning and
transmitted back to the past layer and also
machine learning, enabling individuals to
becomes low. Also, the model's learning
study and practice these methods remotely,
behaviour is too late [22].
often with access to large datasets and
interactive learning modules [18].
2) Convolutional Neural Network (CNN): This
model could be best suited for 2D data. This
5) Transfer Learning: It refers to the technique
network consists of a convolutional filter
where a pre-trained model, usually trained on
for transforming 2D to 3D which is quite
a large dataset of general images like natural
strong in performance and is a rapid
photographs, is adapted and fine-tuned to
learning model. For classification process, it
analyse medical images, allowing for
needs a lot of labelled data. However, CNN
improved accuracy in tasks like disease
faces issues, such as local minima, slow rate
detection, segmentation or anatomical
of convergence, and intense interference by
structure identification [19].
humans [23].
6) Ensemble Learning: It refers to a technique
3) Recurrent Neural Network (RNN): RNNs
where multiple machine learning models
have the ability for recognizing the
are combined to analyze medical images,
sequences. The weights of the neurons are
aiming to achieve better accuracy and
spread through all measures. There are
robustness compared to using a single
many variants such as LSTM, BLSTM,
model alone; essentially, by leveraging the
MDLSTM, and HLSTM. This includes
IJMSRT25MAY080 www.ijmsrt.com 354
DOI: https://doi.org/10.5281/zenodo.15522486
Volume-3,Issue-5,May2025 International Journal of Modern Science and Research Technology
ISSN NO-2584-2706

state-of-threat accuracies in character complete image [28]. The majority of the

recognition, speech recognition, and some object detection techniques based on deep
other natural language processing-related learning continue to depend on convolutional
problems. Learning sequential events can neural networks (CNNs) for classifying pixels
model time conditions [24]. or particles.

4) Registration: Image registration is the

process of merging two or more images to
provide enhanced details, improving image
resolution and assisting medical
practitioners in diagnosis. One approach
proposed a 3D CNN-based architecture for
image identification, focusing on feature
Fig.2. General Architecture of Neural representation. [29] The effectiveness of
Network& Deep Learning [25] this method was evaluated using specific
input datasets. For unsupervised flexible
3.2 Deep Learning Applications: image registration, researchers introduced a
1) Segmentation: Semantic segmentation, a key
hierarchical transformer-based approach,
deep-learning task in medical image analysis, which demonstrated significant
is the process of classifying various areas of an improvements in 3D biomedical imaging,
image into meaningful categories. Computer- particularly in deformation registration.
aided diagnosis is commonly used to extract
anatomical features, detect abnormalities, and
5) Enhancement of Images: Denoising
support medical decision-making. [26]
techniques that utilize deep learning depend
Sophisticated segmentation methods assist in
on data pre-processing to enhance image
biomarker identification, tumour recognition,
quality. A deep generative model,
organ outlining, and lesion segmentation,
constructed with fully connected layers, has
positioning it as one of the most crucial
demonstrated impressive effectiveness even
research areas in medical imaging.
when trained on small datasets, efficiently
extracting information in noisy conditions.
2) Categorization: Computer-assisted diagnosis
During the classification procedure, medical
(CAD) refers to classification in medical
images are examined using convolutional
imaging and is often linked to diagnostic
neural networks (CNNs), frequently
procedures. In classification, one or more
utilizing publicly accessible datasets as
images act as the raw input data, and the result
training benchmarks.
is a particular test factor that defines the
evaluation of the image. [27] Essentially, the
3.3 Deep Learning Frameworks: [30]
classification entails designating one of the
1) Caffe: Caffe is a deep learning library
various specified labels to a particular dataset
specifically developed for Convolutional
by employing pattern recognition methods to
Neural Networks (CNNs). Caffe is
improve precision.
constructed from multiple packages for
computation such as MKL, OpenBLAS,
3) Detection: In medical imaging, detection
and cuBLAS. It contains model fine-tuning,
encompasses the integrated processes of
prediction, and learning tools. It also
classification and localization, concentrating
supports a server application that can be
on recognizing particular objects within an
readily downloaded and accepts both
image. Deep learning is a vital and important
Python and MATLAB APIs.
field of study in computer-aided diagnosis
(CAD). A Key aim of detection is to find and
recognize minor anomalies throughout a
IJMSRT25MAY080 www.ijmsrt.com 355
DOI: https://doi.org/10.5281/zenodo.15522486
Volume-3,Issue-5,May2025 International Journal of Modern Science and Research Technology
ISSN NO-2584-2706

2) TensorFlow: TensorFlow is an open-source improves the diagnosis of breast, lung, and brain
flexible and scalable deep learning cancer, lowering false positives.
framework with wide usage in the
deployment of ML algorithms. It finds its 2) Diabetic Retinopathy Detection: Retinal
use in speech recognition, data analysis, images of diabetic retinopathy can be analysed
nanotechnology, knowledge representation, by using DL algorithms. CNNs detect
and computational linguistics. abnormalities such as haemorrhages and
exudates. AI screening enhances early detection
3) Torch: Torch is a DL library that supports all and ranks high-risk patients first.
types of ML algorithms like multi-layer 3) Diabetes Detection: AI diagnoses diabetes
perceptron, SVMs, Markov models, through retinal scans, MRI, and CT images.
convolutional networks and probabilistic CNNs look for patterns to predict
classification. It supports both CPU and GPUs, complications of diabetes. Feature
and also ported onto Apple, Android, and FPGA- extraction detects early diabetes indicators.
based hardware making it a platform for doing Deep learning systems automate lower
research in deep learning. errors and enhance monitoring.

4) Keras: Keras is a high-level deep learning 4) Cardiac Registration: AI registers cardiac

Python framework designed by François Chollet images from MRI, CT, and
that provides an interface for the construction echocardiography. Deep learning enhances
and training of deep neural networks. It is the diagnosis of heart ailments such as
applied in businesses such as Microsoft, Visa, arrhythmias and CAD. CNNs segment heart
YouTube, Cisco, and Uber. anatomy for accurate functional analysis.

5) Django: Django is not a framework for deep 5) Lung Nodule Classification: AI classifies
learning but a web development framework benign and malignant lung nodules in CT
based on Python. It is utilized in the development scans. CNNs extract shape, texture, and
of secure and scalable web applications. It is density features for classification. Early
involved in deploying AI models within web diagnosis enhances survival using timely
applications but does not directly help in building intervention. Transfer and ensemble
deep learning models. learning boost the accuracy of
classification.
4. MedicalImageAnalysisUsing Deep Learning:
6) Brain Tumour Detection: Deep learning
classifies brain tumours from CT and MRI
scans. CNNs & segmentation networks
differentiate between types of tumours such
as gliomas and metastases. Sophisticated
models enhance the localization of the
tumour with high accuracy.

7) Fetal Localization: AI supports fetal

monitoring through ultrasound image
Fig. 3.Deep learning applications in medical analysis. CNNs identify fetal structures to
mageanalysis [31] evaluate growth and position. Automated
1) Cancer Detection: Deep learning analyses detection recognizes anomalies and
cancer using medical imaging such as X-rays, gestational age.
CTs, and MRI.CNNs segment and classify
tumours, enhancing early detection rates. AI

IJMSRT25MAY080 www.ijmsrt.com 356

DOI: https://doi.org/10.5281/zenodo.15522486
Volume-3,Issue-5,May2025 International Journal of Modern Science and Research Technology
ISSN NO-2584-2706

8) Thyroid Diagnosis: Deep learning identifies algorithms and models that show remarkable
thyroid nodules through ultrasound and CT improvements in diagnostic accuracy.
scan analysis. CNNs classify nodules as
benign or malignant with high accuracy. AI 6. Conclusion:
minimizes unnecessary biopsies and Deep learning has become a versatile tool in
enhances early diagnosis. Automated the analysis of medical images, making it
segmentation allows accurate measurement possible for automated, effective, and very
and feature extraction. accurate disease diagnosis and prognosis.
[39]CNNs and other deep models have shown
9) Prostate Image Segmentation: AI great performance in tasks such as
automatically segments prostate images from segmentation, classification, detection, and
MRI and ultrasound images. CNNs and U- registration, improving medical imaging
Net models enhance tumor localization workflows greatly.
accuracy. Deep learning helps in biopsy Although deep learning promises much for the
planning and cancer treatment. Transfer future of medical imaging, its successful
learning enhances segmentation accuracy on deployment in clinical practice will -depend on
large datasets. the collaboration of AI researchers, clinicians,
and regulatory agencies. Ensuring that deep
learning models are aligned with actual
medical requirements will be crucial to
pushing forward AI-driven healthcare
innovations.

References:
[1] Winsberg F, Elkin M, Macy J, Bordaz V,
Weymouth W. Detection of radiographic
abnormalities in mammograms by means of
optical scanning and computer analysis.
5. Discussion: Radiology. 1967;89:211–5.
Medical Image Processing plays an important role [2] Deep Learning in Medical Image Analysis,
using deep learning methods and specifically Heang-Ping Chan, HHS Public Access, 2021,
convolutional neural networks (CNNs). Research page no. 2
reveals that deep learning methods show better [3] Semmlow JL, Shadagopappan A,
performance in segmentation, classification, [36] Ackerman LV, Hand W, Alcorn FS. A fully
detection, and registration than conventional automated system for screening
machine learning methods. Supervised learning mammograms. Computers and Biomedical
algorithms and methods exhibit great accuracy in Research. 1980;13:350–62. [PubMed:
medical image classification if there are large 7408456]
amounts of annotated data while unsupervised and [4] Deep Learning Application in Medical
self-supervised learning algorithms provide Image Analysis, JUSTIN KER1,
promising performance if labelled data is scarce. 2017,Access.2017.2788044
Deep learning architectures use frameworks such [5] Spiesberger W. Mammogram inspection by
as Tensor Flow, [37] Caffe, Torch, and Keras to computer. IEEE Trans Biomed Eng.
support efficient model training and deployment in 1979;26:213–9. [PubMed: 437802]
medical image analysis. Multiple medical [6] Masayuki Tsuneki1,Deep learning models
treatments such as cancer detection, diabetic in medical image analysis, published by
retinopathy diagnosis, lung nodule classification, Elsevier, 2022, page no. 3
and brain tumor detection use deep learning

IJMSRT25MAY080 www.ijmsrt.com 357

DOI: https://doi.org/10.5281/zenodo.15522486
Volume-3,Issue-5,May2025 International Journal of Modern Science and Research Technology
ISSN NO-2584-2706

[7] Masayuki Tsuneki1,Deep learning models [17] Balasamy K, Suganyadevi S (2021) A

in medical image analysis, published by fuzzy based ROI selection for encryption and
Elsevier, 2022, page no. 3 watermarking in medical image using DWT
[8] Jain G, Mittal D, Takur D, Mittal MK and SVD. Multimedia Tools Appl 80:7167–
(2020) A deep learning approach to detect 7186.
Covid-19 coronavirus with X-ray [18] Plis SM, Hjelm DR, Salakhutdinov R,
images.Elsevier-Biocybern Biomed Eng Allen EA, Bockholt HJ, Long JD, Johnson HJ,
40(4):1391–1405. Paulsen JS, Turner JA, Calhoun VD(2014)
[9] Sethy PK, Behera SK, Ratha PK (2020) Deep learning for neuroimaging: a validation
Detection of coronavirus disease (COVID-19) study. Front Neurosci 8:229.
based on deep features and support vector [19] Sarraf S, Tofighi G (2016) Classification
machine. Int J Math Eng Manag Sci 5(4):643– of Alzheimer’s diseaseusing fMRI data and
651. deep learning convolutional neuralnetworks.
[10] Anavi Y, Kogan I, Gelbart E, Geva O, arxiv: 1603.08631
Greenspan H (2015) A comparative study for [20] Suganyadevi S, Shamia D, Balasamy K
chest radiograph image retrieval using binary (2021) An IoT-based diet monitoring
texture and deep learning classification. In: healthcare system for women. Smart Healthc
Proceedings of the IEEE engineering in Syst Des Secur Priv Asp.
Medicine and Biology Society, pp 2940–2943. [21] Jayashree Moorthy and Usha Devi
[11] Jaiswal AK, Tiwari P, Kumar S, Gupta D, Gandhi; A Survey on Medical Image
Khanna A, Rodrigues JJ (2019) Identifying Segmentation Based on Deep Learning
pneumonia in chest X-rays: a deep learning Techniques. MDPI; 2022, page no. 4
approach. Measurement 145(2):511–518 [22] Tran PV (2016) A fully convolutional
[12] Arevalo J, Gonzalez FA, Pollan R, neural network for cardiac segmentation in
Oliveira JL, Lopez MAG (2016) short axis MRI. arxiv: 1604.00494.
Representation learning for mammography abs/1604.00494
mass lesion classification with convolutional [23] Balasamy K, Ramakrishnan S (2019) An
neural networks. Compute Methods Programs intelligent reversible watermarking system for
Biomed 127:248–25 authenticating medical images using wavelet
[13] Bengio Y (2012) Practical and PSO. Clust Comput 22(2):4431–4442.
recommendations for gradient-based training [24] Yang D, Zhang S, Yan Z, Tan C, Li K,
of deep architectures. Neural networks: tricks Metaxas D (2015) Automated anatomical
of the trade. Lecture Notes in Computer landmark detection on distal femur surface
Science, 7700, Springer, Berlin, Heidelberg. using convolutional neural network. In:
[14] Bengio Y, Simard P, Frasconi P (1994) Proceedings of the IEEE international
Learning long-term dependencies with symposium on biomedical imaging, pg no. 17–
gradient descent is difficult. IEEE Trans 21.
Neural Networks 5(2):157–166. [25] S. Suganyadevi1 · V. Seethalakshmi1 · K.
[15] BenTaieb A, Kawahara J, Hamarneh G Balasamy; A review on deep learning in
(2016) Multi-loss convolutional networks for medical image analysis. Springer Nature.
gland analysis in microscopy. In: Proceedings 2021, Springer Nature, International Journal of
of the IEEE international symposium on Multimedia Information Retrieval (2022):
biomedical imaging, pp 642–645. page no. 24
[16] Havaei M, Guizard N, Chapados N, [26] Jayashree Moorthy and Usha Devi
Bengio Y (2016) HeMIS: hetero-modal image Gandhi; A Survey on Medical Image
segmentation. In: Proceedings of the medical Segmentation Based on Deep Learning
image computing and computer-assisted Techniques. MDPI; 2022, page no. 6
intervention. Lecture Notes in Computer [27] Jayashree Moorthy and Usha Devi
Science, 9901, pp 469–477. Gandhi; A Survey on Medical Image

IJMSRT25MAY080 www.ijmsrt.com 358

DOI: https://doi.org/10.5281/zenodo.15522486
Volume-3,Issue-5,May2025 International Journal of Modern Science and Research Technology
ISSN NO-2584-2706

Segmentation Based on Deep Learning

Techniques. MDPI; 2022, page no. 7
[28] Jayashree Moorthy and Usha Devi
Gandhi; A Survey on Medical Image
Segmentation Based on Deep Learning
Techniques. MDPI; 2022, page no. 8
[29] Jayashree Moorthy and Usha Devi
Gandhi; A Survey on Medical Image
Segmentation Based on Deep Learning
Techniques. MDPI; 2022, page no. 9
[30] Jayashree Moorthy and Usha Devi
Gandhi; A Survey on Medical Image
Segmentation Based on Deep Learning
Techniques. MDPI; 2022, page no. 10
[31] Jayashree Moorthy and Usha Devi
Gandhi; A Survey on Medical Image
Segmentation Based on Deep Learning
Techniques. MDPI; 2022, page no. 6
[32] Saad M. Darwish 1, A New Medical
Analytical Framework for Automated
Detection of MRI Brain Tumor Using
Evolutionary Quantum Inspired Level Set
Technique. MDPI; 2023, page no. 36
[33] Rasha Karam Mahmoud Mohammed;
Acute interstitial pancreatitis;radiopaedia.org;
2024, page no. 14
[34] P. Mohamed Shakeel, M. A. Burhanuddin
& Mohammad Ishak Desa; Automatic lung
cancer detection from CT image using
improved deep neural network and ensemble
classifier; Springer Nature, Volume 34, pages
9579–9592, (2022)
[35] Vahid Tavakoli, Amir A. Amini 1 ; A
survey of shaped-based registration and
segmentation techniques for cardiac images;
sciencedirect; Volume 117, Issue 9, September
2013, Pages 966-989.
[36] Jayashree Moorthy and Usha Devi
Gandhi; A Survey on Medical Image
Segmentation Based on Deep Learning
Techniques. MDPI; 2022, page no. 6
[37] Jayashree Moorthy and Usha Devi
Gandhi; A Survey on Medical Image
Segmentation Based on Deep Learning
Techniques. MDPI; 2022, page no. 9
[38] Chan HP, Hadjiiski LM, Samala
RK.Computer-aided diagnosis in the era of
deep learning. Med Phys 2020;47

IJMSRT25MAY080 www.ijmsrt.com 359

DOI: https://doi.org/10.5281/zenodo.15522486