Journal of Robotics and Control (JRC) Volume 3.
Issue 3.
May 2026 ISSN: 2715-5072.
DOI: 10.
18196/jrc.
Generative Adversarial Networks in Medical Imaging: A Review of Emerging Trends and Insights Yuri Pamungkas 1*.
Myo Min Aung 2.
Muhammad Nur Afnan Uda 3.
Uda Hashim 4 Department of Medical Technology.
Institut Teknologi Sepuluh Nopember.
Surabaya.
Indonesia Department of Mechatronics Engineering.
RMUTT.
Khlong Luang.
Pathum Thani.
Thailand Department of Electronic and Computer Engineering.
Universiti Malaysia Sabah.
Kinabalu.
Malaysia Department of Electrical Electronic Engineering.
Universiti Malaysia Sabah.
Kinabalu.
Malaysia Email: 1* yuri@its.
id, 2 myomin_a@rmutt.
th, 3 nurafnan@ums.
my, 4 uda@ums.
*Corresponding Author AbstractAiGenerative Adversarial Networks (GAN.
are increasingly applied in biomedical imaging for realistic image synthesis, enhancement, and modality translation.
Despite rapid progress, challenges persist in preserving anatomical accuracy, standardizing evaluation, and achieving reliable clinical This review examines recent GAN-based studies, identifying their architectures, applications, datasets, evaluation methods, strengths, and limitations.
The main contribution is a comprehensive synthesis of GAN models across biomedical domains, outlining trends, performance metrics, and research gaps for future exploration.
A structured literature review was conducted, analyzing peer-reviewed studies that met predefined inclusion criteria.
From the screened records, 66 articles satisfied the eligibility criteria for final inclusion.
Data extraction covered GAN types, biomedical applications, dataset characteristics, evaluation metrics, strengths, and limitations.
Applications were categorized, and architectures compared to identify thematic and performance patterns.
Results indicate research concentration in Medical Physics & Reconstruction and Cancer Imaging & Diagnosis.
Advanced models such as StyleGAN.
PGGAN, and CycleGAN deliver superior image realism and enhance downstream performance.
Frequently used datasets include CBIS-DDSM.
ChestX-ray14.
BraTS, and OAI.
Evaluation relies on SSIM.
PSNR, and FID, complemented by clinical metrics such as Dice and AUC.
Strengths include improved image quality, diagnostic accuracy gains via augmentation, and adaptability to diverse imaging modalities.
Limitations involve dataset dependency, anatomical inaccuracies, training instability, high computational demands, and insufficient clinical validation.
In conclusion.
GANs show strong potential for biomedical imaging, but advancements are needed to enhance anatomical fidelity, create standardized evaluation protocols, and address ethical considerations to ensure safe clinical deployment.
KeywordsAiGANs.
Biomedical Imaging.
Image Synthesis.
Deep Learning.
Medical Image Analysis INTRODUCTION The rapid advancements in medical imaging have significantly transformed diagnostic practices, enabling the precise and early detection of a wide array of diseases .
Despite the impressive capabilities of conventional imaging techniques, significant challenges remain in achieving high accuracy and reliability, particularly when distinguishing subtle patterns in complex medical data .
These challenges are further exacerbated by the limitations of available datasets, high variability in image quality, and the need for specialized models that can handle diverse imaging formats .
For example, in fields like oncology, radiology, and neurology, traditional imaging techniques often struggle with issues such as incorrect identifications, low resolution, and challenges in recognizing early-stage abnormalities .
These issues underscore the need for innovative methods that improve the analysis of medical images, particularly in automating interpretation and enhancing diagnostic accuracy .
GANs have emerged as a potentially effective approach to these challenges, particularly in the realm of medical With their unique architecture, comprising a generator and a discriminator.
GANs have shown significant potential in generating realistic medical images, augmenting datasets, and improving image quality .
The key strength of GANs rests in their capacity to generate high-quality artificially generated visuals that closely resemble real medical data, offering a valuable solution in situations where access to large annotated datasets is limited .
GANs have been applied to a variety of applications including enhancing image resolution, reducing noise, partitioning images, and translating across imaging modalities, all of which contribute to enhancing the accuracy and robustness of diagnostic systems .
Over the past few years, the incorporation of GANs within the medical imaging domain has received considerable attention, with numerous studies demonstrating their ability to improve diagnostic outcomes across various applications .
One of the most notable contributions has been in the detection of cancer, where GAN-enhanced types of imaging techniques including mammograms.
CT scans, and MRI scans have shown substantial improvements in classification accuracy .
GANs have also been effectively applied in dermatology, where they help generate synthetic images for skin cancer detection, as well as in radiology and histopathology, where they have enhanced the quality of diagnostic images allowing for better feature extraction and more accurate detection of abnormalities .
These successes underscore the versatility and potential of GANs in diverse medical fields.
The contribution of this review aims to deliver an in-depth examination of the emerging trends in Journal Web site: http://journal.
id/index.
php/jrc Journal Email: jrc@umy.
Journal of Robotics and Control (JRC) ISSN: 2715-5072 GANs for medical imaging, with a particular emphasis on their applications, benefits, along with challenges.
This review will examine various GAN architectures, such as DCGAN.
CycleGAN, and StyleGAN, and their applications in medical imaging, including data augmentation, image reconstruction, and modality translation.
We will also discuss the evaluation metrics frequently applied within GAN-based clinical imaging research, such as SSIM.
PSNR, and FID, and how these indicators serve to evaluate the quality and applicability of generated visual outputs in clinical settings.
Furthermore, this review will highlight the novel contributions of GAN-based techniques in overcoming the limitations of traditional medical imaging approaches.
Specifically, we will discuss how GANs can produce artificial datasets of superior quality, offering a remedy for the challenge of limited data availability in healthcare Additionally, we will explore the potential for GANs to handle various imaging modalities and the possibility of creating generalized models that can be applied across different medical disciplines.
Through this analysis, we seek to provide a detailed understanding of the present forefront of GAN applications in clinical imaging, with emphasis on obstacles, advancements, and forthcoming directions in this rapidly evolving field.
II.
METHODOLOGY
Eligibility Criteria To ensure a focused and comprehensive review of the latest developments, this study adopted a clearly defined set of eligibility criteria for selecting relevant literature.
Articles were sourced exclusively from the ScienceDirect.
Ie Xplore.
SpringerLink, and Wiley databases, which were chosen for its extensive and reliable coverage of peerreviewed publications in the fields of medical technology and artificial intelligence.
The search strategy was based on the use of specific keywords related to GANs and their applications in medical imaging.
These keywords included terms such as AuGenerative Adversarial NetworkAy.
AuGANAy.
Aumedical imagingAy, and additional terms like AutrendsAy.
AuapplicationsAy, and Aucurrent insightsAy to target publications focused on the use of GANs in medical imaging contexts.
The search was restricted to publications from the past three years .
2Ae2.
to ensure the inclusion of recent advancements and emerging trends in the field.
Additional inclusion criteria required that articles be available in full text, written in English, and directly related to the biomedical applications of GANs in imaging modalities such as radiology, histopathology, or other diagnostic domains.
This review included only original research articles that presented technical implementations, methodological advancements, or clinical implications of GAN-based approaches in medical imaging.
Review articles, editorials, conference abstracts, and non-peer-reviewed materials were excluded.
Studies were also excluded if they focused on non-medical domains, did not employ GANbased techniques, or lacked sufficient methodological or experimental detail.
This rigorous filtering process was intended to ensure that the final set of included studies provided robust empirical evidence and meaningful insights into the evolving role of GANs in medical imaging.
Information Sources The ScienceDirect.
Ie Xplore.
SpringerLink, and Wiley databases were electronically searched for eligible The search was conducted for publications published between January 2022 and December 2024, focusing on articles that utilized specific keywords related to GANs and their applications in medical imaging.
Search Strategy The search for relevant studies was conducted using a carefully crafted search query to ensure the identification of comprehensive and up-to-date literature.
The query employed was ("Generative Adversarial Network" OR "GAN") AND ("medical imaging") AND ("trends" OR "applications" OR "current insights").
This search string was specifically designed to capture a wide array of research articles that focus on the use of Generative Adversarial Networks (GAN.
within the domain of medical imaging.
The terms "Generative Adversarial Network" and "GAN" were included to ensure that the search encompassed all publications related to GAN technologies, including those using the abbreviation AuGANAy.
By incorporating the keyword "medical imaging", the search was further refined to focus on studies that apply GANs in various diagnostic imaging techniques, such as radiology, histopathology, and other medical imaging fields.
To ensure the query captured the most current and relevant studies, additional keywords like AutrendsAy.
AuapplicationsAy, and Aucurrent insightsAy were used.
These terms were specifically chosen to target publications that explore the evolving and cutting-edge uses of GANs, ensuring that only studies that present current advancements, emerging applications, and state-of-the-art insights in GAN-based medical imaging were included.
This strategic approach in the search query helped narrow down the articles to those that not only employed GANs in medical imaging but also discussed new trends and innovations within the field, ensuring the inclusion of the most relevant and recent research on the topic.
Study Selection and Quality Assessment The procedure for selecting studies incorporated in this review followed a structured approach to guarantee that only pertinent and rigorous research was taken into account.
The first stage involved assessing the title of each identified Articles were screened for their relevance to the topic of GANs within clinical imaging.
If the title clearly indicated a focus on the application, methodology, or impact of GANs within clinical imaging, it proceeded to the next step in the selection process.
If the title appeared unrelated to the subject or lacked clarity regarding the use of GANs, the article was excluded from further consideration.
Following the title assessment, the next step was to review the abstract and keywords of each article.
The abstract provided an initial overview of the study's objectives, methods, and conclusions.
During this phase, particular attention was given to ensuring that the study specifically addressed the application of GANs in medical imaging.
Articles whose abstracts mentioned GANs applied to imaging modalities such as radiology, histopathology, or other Yuri Pamungkas.
Generative Adversarial Networks in Medical Imaging: A Review of Emerging Trends and Insights Journal of Robotics and Control (JRC) ISSN: 2715-5072 diagnostic imaging fields were included.
Keywords were also evaluated to confirm that they were aligned with the scope of the review, focusing on relevant terms such as AuGANAy.
Aumedical imagingAy.
AuapplicationsAy, and AutrendsAy.
If the abstract and keywords suggested a mismatch with the review's scope, the article was excluded.
The final phase in the study selection involved analyzing the full text of the articles that passed the previous stages.
This step was critical to assess the depth of the studyAos methodology, experimental design, and overall relevance.
Only original research articles that employed GAN-based methods and provided sufficient empirical data were included in the review.
Articles were evaluated for their methodological rigor, including the clarity of the study design, the reliability of the findings, and the soundness of the evaluation.
Studies lacking detailed methodology, statistical analysis, or those not contributing substantial insights into the uses of GANs within clinical imaging were excluded at this stage.
By thoroughly analyzing the full text, the review ensured that only studies with a strong scientific foundation and relevance to the research question were selected to be incorporated.
Fig.
PRISMA flow diagram of study selection.
TABLE I.
SELECTED PAPERS ACCORDING TO THE SPECIFIED CRITERIA
Author & Year Yamazaki et al.
Karar et al.
, 2022
Zhang et al.
, 2022
Semi-Supervised GAN.
AC-GAN
SOUP-GAN
Biomedical Application Mammogram Synthesis.
Breast Cancer COVID-19 Diagnosis (Ultrasoun.
MRI Super-resolution .
Toda et al.
, 2022
Liu et al.
, 2022
StyleGAN
SSGAN
Lung Cancer.
Tumor Imaging Chest X-ray.
Histopathology .
Ali et al.
, 2022
3D CycleGAN
Skeletal Muscle Contraction Dynamics Ref
GAN Type
CR-GAN
Dataset CBIS-DDSM.
INBreast.
CMMD
Public Lung US POCUS Dataset T1 MRI.
T2-FLAIR MRI,
CT .
ChestX-ray14.
BreakHis Simulated Experimental Evaluation Method PSNR.
SSIM.
FID.
MS-SSIM,
Cos_sim Accuracy.
Precision.
Recall.
Specificity.
F1-score PSNR.
SSIM.
RMSE.
FID,
Perceptual Loss PSNR.
SSIM.
FID.
LPIPS
Accuracy.
F1-Score.
AUC.
Sensitivity.
Specificity DBhat.
Correlation.
SSIM.
Temporal Consistency Yuri Pamungkas.
Generative Adversarial Networks in Medical Imaging: A Review of Emerging Trends and Insights Journal of Robotics and Control (JRC) ISSN: 2715-5072 Ref Author & Year Khaled et al.
, 2022
GAN Type Multi-stage GAN Biomedical Application Brain Segmentation .
Guo et al.
, 2022
X-ray Tomography .
Qu et al.
, 2022
Physics-assisted GAN cGAN Sun et al.
, 2022
HA-GAN
Gan et al.
, 2022
HieGAN Liang et al.
, 2022
Ad CycleGAN Shah et al.
, 2022
cGAN Salvia et al.
, 2022
DCGAN
Waqas et al.
, 2022
PGGAN
Salini et al.
, 2022
CycleGAN Tai et al.
, 2022
rAC-GAN Rahman et al.
, 2022
cGAN Singla et al.
, 2022
Velzen et al.
, 2022
Dataset MICCAI iSEG.
MRBrainS Synthetic CircuitFaker Evaluation Method Dice Coefficient.
Computational Time.
Accuracy BER.
SSIM.
MSE.
FID Collaborative Learning.
Federated Learning 3D Image Synthesis (CT.
MRI) Knee Imaging.
Osteoarthritis Diabetic Retinopathy.
Bone Age COPDGene.
GSP Accuracy.
AUC.
MAE.
Dice OAI dataset COVID-19 X-ray Image Synthesis Forearm Vein Segmentation Kaggle COVID-19 Hyperspectral Skin Cancer Dataset OAI dataset cGAN CycleGAN Skin Cancer Diagnosis (Hyperspectra.
Knee MR Imaging.
Data Augmentation Retinal Fundus Image Generation Lung Cancer.
Pulmonary Embolism Dose Mapping.
Proton Therapy Pleural Effusion Diagnosis Cardiac Calcium Scoring AM Score.
Mode Score.
MAE.
WD.
FID.
IS MSE.
RMSE.
PSNR.
FID.
UIQI.
VIF.
Accuracy Sensitivity.
Specificity.
Accuracy.
Dice.
AUC FID.
Accuracy.
Precision.
Recall.
AM Score.
Mode Score.
Dice Alrashedy et al.
Campello et al.
Kelkar et al.
, 2023
Chaudhury et al.
Li et al.
, 2023
DCGAN
Brain Tumor Detection cGAN Cardiac MRI Aging Kaggle Brain Tumor Dataset UK Biobank StyleGAN2 Progressive GAN Medical Image Synthesis Breast Cancer (Ultrasoun.
Simulated datasets BUSI Dataset cGAN OCT Super-Resolution FGAN Stroke Screening (Facial Weaknes.
Endoscopy.
Gastric Cancer Phienphanich et al.
Magalhyes et al.
AlTakrouri et al.
Cackowski et al.
Kou et al.
, 2023
Coronary.
Fish Corneal.
Rat Retina Real Subjects dataset Campos et al.
, 2023
cGAN Tolpadi et al.
, 2023
PatchGAN Ikuta et al.
, 2023
TextureWGAN Cycle-GAN Kalantar et al.
Buczek et al.
, 2023
Luong et al.
, 2024
DCGAN
Kumaar et al.
, 2024
DCGAN, cGAN
ISRGAN
NIR dataset
DRIVE
Yunnan & Chongqing Custom (Geant4/GATE MIMIC-CXR RTP CT.
NLST
Public endoscopic images FID.
MMD.
IS, t-test.
F1.
SSIM.
AUC.
Accuracy.
Precision AUC.
Sensitivity.
Specificity.
Accuracy.
MRE.
CV.
FID.
Classifier Consistency ICC.
CAC pseudo-mass.
Agatston Accuracy.
Precision.
Recall.
AUC.
Loss MAE.
FID.
PSNR
FID.
JS Divergence.
PCA
Accuracy.
Precision.
Recall.
F1.
AUC
PSNR.
SSIM.
SFD.
EPI
AUC.
Sensitivity.
Specificity.
F1.
Confusion Matrix Accuracy.
Sensitivity.
Specificity.
F1.
AUC
PSNR.
SSIM.
MAE.
MSE.
FID
Image Super-Resolution (Diagnostic.
MRI Image Harmonization CIFAR-10 DOF Extension.
Image Fusion Food Safety (Poultry Defect Detectio.
Rheumatoid Arthritis Imaging CT Reconstruction.
Denoising.
Super-Resolution COVID-19 Pneumonia (CT) Multi-focus dataset CXR.
NCT-CRC-HE-100K
BUSI Dataset Accuracy.
Precision.
Recall.
F1 Style-based GAN Chest X-ray.
Colorectal Cancer Imaging Breast Cancer Classification (Ultrasoun.
Brain Tumor Classification AUC.
Sensitivity.
Specificity.
FPR.
FleissAo Kappa Accuracy.
Precision.
Recall.
F1 T1 MRI dataset Mohammed et al.
Lerch et al.
, 2024
cGAN Fabric Defect Detection AITEX Fabric Dataset cGAN Breast Ultrasound BUSI Dataset Alqushaibi et al.
Di Giammarco et
, 2024
cGAN Colon Cancer Segmentation Kvasir-SEG Accuracy.
Precision.
Recall.
F1.
ROC-AUC.
Confusion Matrix Accuracy.
Precision.
Recall.
F1.
Loss.
ROC-AUC
Balanced Accuracy.
ECE.
CNR, gCNR Dice.
IoU.
F-beta.
SSIM.
MAE
DCGAN
Ophthalmology (Retinal Fundu.
MedMNIST
VAE-GAN
MPGAN
cGAN ABIDE.
OASIS.
SRPBS
Private (SWIR) 27 Rheumatoid Arthritis Patients dataset
LIDC/IDRI
NCCID
SSIM.
AUC, t-SNE.
Confusion Matrix PSNR.
SSIM.
MTF.
Perceptual Loss Precision.
Recall.
F1.
BACC.
Visual inspection nRMSE.
PSNR.
SSIM.
Occlusion Maps PSNR.
SSIM.
Texture Analysis Accuracy.
Precision.
Recall.
F1.
AUC.
Loss Yuri Pamungkas.
Generative Adversarial Networks in Medical Imaging: A Review of Emerging Trends and Insights Journal of Robotics and Control (JRC) ISSN: 2715-5072
Ref
Author & Year
Al-Adwan, 2024
GAN Type
DCGAN
Biomedical Application Brain Tumor MRI Dataset Brain Tumor MRI Dataset CycleGAN Echocardiography EchoNet CycleGAN COVID-19 CT-Scan Dice.
Sensitivity.
Specificity CycleGAN Neuroimaging (Schizophreni.
Holographic Microscopy Mosmed.
Coronacases.
Radiopedia COBRE.
ABIDE
Rakhmetulayeva et
, 2024
Elmourabit et al.
Yamaguchi et al.
Galande et al.
, 2024
Zhou et al.
, 2024
ProGAN SSIM.
MSE.
Phase-SNR.
Visual Inspection ROC.
AUC.
PSNR.
Phase-SNR Dey et al.
, 2024
cGAN Hymon et al.
, 2024
cGAN Xu et al.
, 2024
SR-GAN
WGAN-GP
cGAN Zhang et al.
, 2024
Hamghalam et al.
Irtaza et al.
, 2024
Signal Detection.
Imaging Systems Optimization Polarimetric Imaging (Target Detectio.
Radiotherapy (Multimodal Registratio.
Infrared Imaging.
SuperResolution Low-Dose CT Denoising Brain Tumor Segmentation Cervical cells.
RBCs.
ISBI
Lumpy Object Model.
ADNI MRI
Polarimetric Dataset
DCGAN
Lung Disease Classification NIH Chest X-ray 14 Freitas et al.
, 2024
cGAN Private WLC video cGAN Thandiackal et al.
Rai et al.
, 2024
Bladder Cancer Segmentation Histopathology Classification .
Han et al.
, 2024
RCGAN
Breast Cancer Detection (Ultrasoun.
Pneumonia Diagnosis (CXR) BrEaST.
BUSI.
Thammasat.
HMSS
Guangzhou CXR dataset cGAN Brain Extraction (MRI) .
Moazami et al.
Oliveira et al.
, 2024
Long et al.
, 2024
StyleGAN2ADA cGAN StyleGAN-XL TopoSinGAN Kokomoto et al.
Ahmadkhani et al.
Cui et al.
, 2024
Retinal Imaging (AMD Detectio.
Ultrasound CT.
Breast Cancer Imaging Pediatric Dentistry (Growth Predictio.
Dental Growth Prediction cGAN MRI Segmentation .
Song et al.
, 2024
cGAN Yang et al.
, 2024
SCGAN
Thyroid Disease (Ultrasound Elastograph.
Fundus Imaging.
Ophthalmology NFBS.
CC359.
LPBA.
IBSR iChallenge-AMD.
ODIR2019.
RIADD.
STARE Synthetic & Phantom 8,092 panoramic 8,092 panoramic BUS2017.
DDTI.
LiTS.
ATLAS.
BRATS 2015 Shanghai Sixth PeopleAos Hospital dataset Custom & Public datasets .
Li et al.
, 2024
CycleGAN
HDGAN
StyleGAN3 CT Radiology.
Distribution Shift
RESULTS AND DISCUSSIONS
Generative Adversarial Networks in Medical Imaging In recent years.
GANs have become an integral part of medical imaging workflows, addressing challenges such as limited datasets, variability in image acquisition, and the need for highAaquality, realistic data.
Within the domain of data generation.
GAN variants have been applied to produce clinically relevant synthetic images across a wide range of modalities and applications.
Examples include CRAaGAN for mammogram synthesis in breast cancer screening .
StyleGAN and its derivatives such as StyleGAN2.
StyleGAN3.
ADA, and XL for lesionAaspecific image synthesis and developmental growth modeling in ultrasound.
Evaluation Method PSNR.
MSE.
SSIM.
Accuracy.
Precision.
Recall.
PSNR.
SSIM.
FID.
User Study VBM.
Age Prediction.
Dice Prostate Cancer Dataset PSNR.
SSIM.
Precision.
Recall, mAP.
Dice.
MAE
Custom infrared dataset PSNR.
SSIM.
FID
AAPM-Mayo LDCT BraTS'13.
BraTS'18 PSNR.
SSIM.
FID.
L1 loss Dice.
Sensitivity.
PPV.
Hausdorff, p-values Precision.
Recall.
F1.
Binary Accuracy.
AUC
mAP, mAP50, mAP75.
Precision.
Recall F1, mAP, mAP50, mAP75 K-16.
K-19.
CRC-TP
IMR & YA-CT images Accuracy.
Precision.
Recall.
F1.
TTest.
Wilcoxon Accuracy.
Precision.
Sensitivity.
Specificity.
F1.
AUC
Dice.
Sensitivity.
PPV.
FID.
Sensitivity.
Specificity.
Accuracy.
Expert Eval PSNR.
SSIM, nRMSE FID.
Visual inspection.
Latent FID.
Visual inspection.
Latent Dice.
IoU.
Precision.
Recall PSNR.
SSIM.
MSE.
Rago Score.
Perceptual Study BRISQUE.
PIQE.
SSIM.
PSNR, uscore.
F1.
Sensitivity.
Specificity.
AUC
SSIM.
MS-SSIM.
T-SNE,
Segmentation accuracy CT, and panoramic radiographs .
,34,67,70-.
, and PGGAN or ProGAN for progressively refined image detail .
HighAaresolution designs such as HDGAN .
and MPGAN .
have been used for specialized imaging tasks, while 3D and hierarchical attention architectures such as Ha GAN .
and HieGAN .
have generated volumetric data with improved structural realism for CT.
MRI, and knee osteoarthritis imaging.
CycleAaconsistent approaches have enabled unpaired imageAatoAaimage translation for retinal fundus images .
, cardiac CT calcium scoring .
COVIDAa19 chest CT adaptation .
, echocardiography enhancement .
, and schizophrenia neuroimaging harmonization .
More recently, domain shift robustness Yuri Pamungkas.
Generative Adversarial Networks in Medical Imaging: A Review of Emerging Trends and Insights Journal of Robotics and Control (JRC) ISSN: 2715-5072 has been explicitly tested in CT radiology to ensure model generalization beyond the training environment .
A significant proportion of research focuses on reconstruction, superAaresolution, and denoising, moving beyond perceptual quality toward diagnostic sufficiency.
MRI superAaresolution has been advanced with SOUPAaGAN .
, which incorporates perceptual constraints, while OCT imaging has benefited from cGAN designs with edgeAa sensitive metrics .
SuperAaresolution in diagnostic imaging has been addressed by ISRGAN .
and SRAaGAN .
, while lowAadose CT denoising has achieved stability and high fidelity through WGANAaGP with L1 loss regularization .
Texture preservation, a key diagnostic requirement often overlooked by conventional perceptual metrics, has been addressed with TextureWGAN .
Physicsassisted GANs .
integrate acquisition priors into the generative process for XAaray tomography, and ProGAN has been used for optimizing detectability in imaging systems .
, representing a shift toward physicsAainformed, applicationAa aware image synthesis.
In segmentation, registration, and quantitative mapping.
GANAabased frameworks have improved spatial delineation and crossAamodality alignment.
MultiAastage GANs have been used for infant brain MR segmentation .
, while NIR forearm vein segmentation .
, colon cancer segmentation .
, and bladder cancer detection in cystoscopic video .
have leveraged conditional GANs for precise boundary Brain tumor segmentation using BraTS datasets has been improved with cGAN architectures .
, and prostate radiotherapy workflows have incorporated GANAa based multimodal registration .
CycleGAN has supported domain adaptation in calcium scoring .
COVIDAa19 CT harmonization .
, echocardiography enhancement with expertAavalidated outputs .
, and neuroimaging adaptation for schizophrenia research .
These studies illustrate that GANs are not merely image generators but also facilitators of accurate clinical measurements, improving indices such as Agatston scores and Dice coefficients even when labeled data is scarce.
For classification.
GANs have been embedded into pipelines to balance class distributions and improve performance in dataconstrained scenarios.
SSGAN has enhanced chest XAaray and histopathology classification .
, cGANAabased augmentation has supported pleural effusion detection .
, and DCGAN or RCGAN augmentation has improved lung disease .
and pneumonia detection .
StyleGAN and DCGAN have been applied to breast ultrasound classification .
,47,50,.
and brain tumor MRI classification .
, with some studies incorporating deploymentAaoriented metrics such as Expected Calibration Error (ECE) and contrastAatoAanoise ratio (CNR/gCNR) .
Applications extend beyond traditional medical tasks to include gastric cancer detection in endoscopy .
and defect detection in safetyAacritical contexts such as fabric inspection .
and polarimetric imaging .
, demonstrating the methodological adaptability of GAN frameworks.
From a methodological standpoint, the reviewed works span multiple GAN families, such as conditional GANs .
GAN, cWGAN) for labelAaguided synthesis and mapping, cycleAaconsistent GANs .
D, 3D, and 3DAapatc.
for unpaired domain translation, styleAabased architectures for highAa resolution and latentAaspace manipulation, progressive growth texture/attention/hierarchical approaches to protect fine anatomical structures .
Increasingly, volumetric fidelity is addressed through 3D synthesis and patchAabased processing .
,21,.
, and privacyAapreserving federated training is explored to address multiAasite heterogeneity .
These developments signal a transition from purely data driven generative modeling toward clinically robust, generalizable architectures.
The dataset ecosystem reflects both diversity and clinical relevance, encompassing mammography (CBISAaDDSM.
INBreast.
CMMD) .
, chest radiography (ChestXAaray14.
NIH CXR14.
NCCID.
MosMe.
,45,55,64,.
, largeAa scale multiAasite MRI and CT cohorts (ABIDE.
OASIS.
SRPBS.
LIDC/IDRI.
AAPMAaMay.
,44,.
, ultrasound datasets such as BUSI and multiAacenter breast image collections .
,47,50,.
, ophthalmic datasets including DRIVE, iChallengeaMD.
ODIR, and RIADD .
,69,.
, osteoarthritis OAI knees .
, cardiac UK Biobank and EchoNet .
, and specialized collections for radiotherapy, urology, and elastography .
,60,65,.
Private or simulated datasets are also common in domains where annotated data is scarce, such as CT tumors, proton dose mapping, infrared, and polarimetric imaging .
,29,39,58,.
, underscoring persistent challenges in open data availability.
Evaluation practices have evolved into multiAaobjective frameworks that combine fidelity, perceptual, and clinical task metrics .
Common measures include PSNR and SSIM .
ith variants like MSAaSSIM.
UIQI, and VIF), perceptual scores such as FID and LPIPS, and statistical distribution tests like MMD or PCA divergence.
TaskAa specific performance is evaluated using Dice.
IoU, and Hausdorff distance for segmentation and registration, and AUC, sensitivity, specificity, and F1Aascore for classification.
Texture and phaseAasensitive metrics such as PhaseAaSNR, texture analysis, and EPI/SFD are used where structural detail is critical .
,44,57,.
Importantly, some studies incorporate expert reader studies or clinical user evaluations .
,69,.
, and a few report calibration metrics such as ECE .
, reflecting a growing focus on deployment readiness.
Across 2022-2024, the research trend moves decisively from data augmentation toward deploymentAaoriented generative pipelines.
Key developments include federated learning for privacy and crossAasite robustness .
, harmonization of multiAasite MRI data .
, explicit distribution shift evaluation .
, textureAapreserving SR and denoising .
, and the integration of physicsAabased priors .
,29,.
Persistent gaps remain, including limited external validation across diverse healthcare settings, underreporting of calibration and uncertainty, reliance on nonAapublic datasets, and occasional misalignment between perceptual and clinical outcome metrics.
Promising directions involve scaling 3D generative modeling for wholeAaorgan and temporal analysis .
,21,.
, leveraging styleAaspace control for longitudinal predictions and developmental studies .
, refining physicsAainformed architectures for safetyAa Yuri Pamungkas.
Generative Adversarial Networks in Medical Imaging: A Review of Emerging Trends and Insights Journal of Robotics and Control (JRC) ISSN: 2715-5072 critical tasks .
,58,.
, and developing endAatoAaend evaluation pipelines that link image quality directly to clinical decision performance .
,40,50,.
Taken together, these studies illustrate how GANs have matured from tools for synthetic data generation into clinically relevant, integrated components of medical imaging systems, enhancing reconstruction, superAaresolution, segmentation, and classification while moving toward the rigor and reliability required for realAaworld healthcare applications.
GAN Type GANs are a category of advanced learning algorithms designed to produce artificial information that closely mirrors actual-world examples in statistical terms .
The architecture is built upon two neural network components:
the Generator (G), which produces fabricated information, and the Discriminator (D), which assesses the genuineness of the provided input.
These two models are trained in an adversarial framework where the Generator attempts to generate results that imitate the distribution of actual data, while the Discriminator acquires the ability to differentiate authentic samples from artificially created counterparts .
Different variants of GANs have been developed and tailored to meet specific clinical objectives, ranging from superresolution and denoising to segmentation, synthesis, and classification support.
The selection of a GAN framework is frequently influenced by the type of imaging technique, the intended clinical use, and the preferred trade-off between accuracy and processing efficiency .
The following presents the distribution of GAN types most frequently employed in biomedical contexts, with a particular focus on medical imaging.
TABLE II.
DISTRIBUTION OF GAN TYPES
GAN Type Frequency References .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
Conditional GAN
GAN) CycleGAN
StyleGAN
DCGAN
Wasserstein GAN
(WGAN)
, .
Modified GAN , .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
The distribution of GAN types in medical imaging research .
shows a clear preference for Conditional GANs .
GAN.
, which account for 25 studies.
This dominance reflects the versatility of cGANs in incorporating label or attribute information during training, enabling more targeted synthesis, domain translation, and data augmentation for supervised and semiAasupervised tasks.
Their frequent use in applications such as segmentation, classification, and modality translation underlines their adaptability across diverse imaging modalities and clinical objectives.
The second most frequent group is Modified GANs, appearing in 20 studies.
This category includes various architecture adaptations, such as physicsassisted GANs, hierarchical attention GANs, multiAastage GANs, and textureAa preserving GANs.
The high number indicates a strong trend toward tailoring baseline GAN architectures to meet the specific demands of medical imaging, whether by integrating domain knowledge .
, physicsAabased constraint.
or optimizing for volumetric data and highAaresolution outputs.
This also highlights that researchers often find it necessary to go beyond standard architectures to achieve clinically acceptable results.
CycleGANs appear in 7 studies, reflecting their established utility in unpaired imageAatoAaimage translation.
Their ability to align domains without paired datasets is particularly valuable in medical imaging, where paired multiAa modal or crossAavendor data are rare.
Applications include domain adaptation between imaging modalities and harmonization across scanners or institutions.
StyleGANs are used in 6 studies, mainly for highAaquality, highAaresolution synthesis and latentAaspace manipulation.
Their generative control and visual fidelity make them attractive for tasks requiring fine anatomical detail preservation, such as lesion synthesis or developmental modeling.
While StyleGANAos adoption is lower than cGAN, its unique strengths are increasingly recognized for specialized tasks.
DCGANs are found in 7 studies, mostly as a baseline architecture for data augmentation and classification support.
Although DCGAN is one of the earliest GAN variants, its simplicity and relative ease of training still make it useful for straightforward augmentation scenarios or as a foundation for hybrid designs.
Finally.
Wasserstein GANs (WGAN.
, including WGANAaGP, appear in only 2 studies.
This low frequency may be due to their more specialized use cases, such as improving training stability in superAaresolution and denoising tasks.
While WGANs offer theoretical advantages in training dynamics, they are often embedded within broader, more specialized architectures rather than used in their pure form.
Biomedical Application GANs have emerged as a transformative technology in biomedical imaging, enabling innovations in image synthesis, enhancement, and analysis across diverse medical domains .
By leveraging their ability to generate highfidelity, realistic images.
GANs address long-standing challenges such as limited annotated datasets, variations in imaging protocols, and the need for modality translation .
Their applications span from improving image quality in lowdose or time-constrained acquisitions to supporting disease diagnosis through synthetic data augmentation and crossdomain adaptation.
Table 3 summarizes the distribution of GAN applications across various biomedical imaging categories, highlighting both well-established use cases and emerging research areas.
TABLE i.
GAN IN BIOMEDICAL APPLICATIONS
Biomedical Application Category Cancer Imaging & Diagnosis Frequency References .
, .
, .
, .
, .
, .
, .
, .
, .
, .
Yuri Pamungkas.
Generative Adversarial Networks in Medical Imaging: A Review of Emerging Trends and Insights Journal of Robotics and Control (JRC) Biomedical Application Category Cardiac Imaging Neurological Imaging Ophthalmology Musculoskeletal Pulmonary & Thoracic Imaging Vascular Imaging Dental Medical Physics & Reconstruction Infectious Diseases Histopathology Other Biomedical Applications ISSN: 2715-5072 Frequency References .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
, .
Based on Table 3, the utilization of GANs in biomedical imaging can be grouped into several major categories, each reflecting distinct clinical and research priorities.
The highest representation is found in Medical Physics & Reconstruction .
, where GANs are employed to enhance image quality through super-resolution, denoising, image harmonization, and reconstruction.
This dominant share indicates that improving the fidelity and diagnostic utility of medical images is one of the most impactful contributions of GANs.
These applications are critical in scenarios with lowquality or noisy data, such as low-dose CT scans.
MRI harmonization across different scanners, and multimodal image registration.
GAN-based approaches in this category help reduce acquisition time, minimize radiation exposure, and improve interoperability between imaging systems.
Cancer Imaging & Diagnosis .
forms the second-largest category, demonstrating GANsAo crucial role in oncological applications.
Their use spans data augmentation for training classifiers, tumor segmentation, and synthesis of realistic cancer-related imaging for modalities such as mammography.
CT.
MRI, ultrasound, and endoscopy.
These applications are essential in addressing class imbalance, increasing dataset diversity, and improving the robustness of AI diagnostic tools across various cancer types, including breast, lung, brain, skin, gastric, colorectal, and bladder A moderate level of research activity is observed in Neurological Imaging .
Ophthalmology .
Musculoskeletal .
, and Pulmonary & Thoracic Imaging .
In neurological applications.
GANs support brain segmentation, lesion detection, and neuroimaging harmonization, while in ophthalmology they facilitate retinal image synthesis, enhancement, and disease detection such as age-related macular degeneration (AMD).
Musculoskeletal uses include knee osteoarthritis imaging and skeletal muscle modeling, whereas pulmonary applications involve chest X-ray enhancement, lung disease classification, and pneumonia diagnosis.
Other specific categories show lower frequencies but notable potential.
Cardiac Imaging .
employs GANs for calcium scoring.
MRI analysis, and echocardiography enhancement.
Dental applications .
involve pediatric growth prediction and dental morphology modeling.
Vascular Imaging .
focuses on vein segmentation in near-infrared (NIR) imaging.
Infectious Diseases .
target COVID-19 diagnosis and domain adaptation for related modalities.
Histopathology .
highlights the underexplored but highly relevant area of applying GANs for pathology slide classification and Other Biomedical Applications .
capture diverse but less common use cases, including signal detection optimization, polarimetric imaging, and safetycritical inspection tasks .
, fabric defect detection in biomedical manufacturing context.
The distribution indicates that GAN research in biomedical imaging is currently concentrated in two main domains, image quality enhancement (Medical Physics & Reconstructio.
and cancer-related applications.
This focus reflects both the immediate clinical impact and the feasibility of GAN integration into these areas due to available datasets and well-defined evaluation metrics.
However, lower representation in areas like histopathology, vascular imaging, and infectious diseases suggests significant research gaps and opportunities for innovation.
Expanding GAN applications into these underexplored domains could help address emerging clinical challenges, especially where data scarcity, cross-domain variability, or the need for high precision is Furthermore, the balance between diagnostic accuracy, image realism, and clinical interpretability will remain central to GAN adoption in healthcare settings.
Dataset The datasets used across the reviewed studies on GANs in biomedical imaging demonstrate a combination of widely recognized public repositories and institution-specific private Public datasets such as CBIS-DDSM.
INBreast, and CMMD are frequently used for breast cancer mammography synthesis due to their standardized imaging protocols and rich annotation quality, which make them ideal for evaluating GAN-based image generation.
Similarly, large-scale open datasets like ChestX-ray14.
NIH Chest Xray 14, and MIMIC-CXR provide diverse chest radiographs for training GAN models in pulmonary and thoracic imaging tasks, such as pneumonia detection and lung disease These repositories offer the advantage of size and variety, but often require careful preprocessing due to heterogeneous quality and labeling standards.
In the neuroimaging domain, datasets such as MICCAI iSEG.
MRBrainS 2013.
BraTS .
3 and 2018 edition.
OASIS, and ABIDE are used extensively for segmentation, tumor classification, and MRI harmonization tasks.
These datasets are valuable for their multi-modal imaging formats .
T1.
T2.
FLAIR) and standardized challenges, which allow for objective benchmarking of GAN architectures.
Moreover, studies focusing on cross-modality synthesis or harmonization often use combinations of MRI and CT datasets, such as COPDGene and GSP, to evaluate the ability of GANs to maintain anatomical fidelity while translating between modalities.
Ophthalmology applications frequently utilize datasets like DRIVE.
MedMNIST.
ODIR-2019, iChallenge-AMD, and STARE for retinal fundus image generation, enhancement, and disease detection.
The relatively high quality and specific disease annotations in these datasets make them suitable for GAN-driven diagnostic Yuri Pamungkas.
Generative Adversarial Networks in Medical Imaging: A Review of Emerging Trends and Insights Journal of Robotics and Control (JRC) ISSN: 2715-5072 support, particularly for age-related macular degeneration (AMD) and diabetic retinopathy.
Likewise, gastrointestinal cancer detection and segmentation often rely on datasets such as NCT-CRC-HE-100K and Kvasir-SEG, which provide histopathology or endoscopic images with pixel-level A notable proportion of studies rely on private or proprietary datasets, particularly in cases where public data are scarce or the imaging modality is highly specialized.
Examples include near-infrared (NIR) forearm vein segmentation datasets, synthetic and phantom datasets for ultrasound CT imaging, and institution-specific collections for dental growth prediction, thyroid elastography, and bladder cancer segmentation.
While private datasets offer highly controlled imaging conditions, they may limit reproducibility and cross-institutional validation unless shared through collaborative agreements.
Finally, synthetic datasets and simulated environments also play a role, particularly in physics-based GAN applications and imaging systems optimization.
Simulated data generated from platforms like Geant4/GATE or synthetic phantoms allow researchers to explore GAN performance in controlled scenarios without patient privacy concerns.
However, translating results from synthetic to real-world clinical data remains a challenge due to differences in texture, noise patterns, and anatomical variability.
TABLE IV.
DATASET GROUPING
Ref Dataset CBIS-DDSM.
INBreast.
CMMD
Public Lung US POCUS Dataset T1 MRI.
T2-FLAIR MRI.
CT .
ChestX-ray14.
BreakHis .
Simulated Experimental .
MICCAI iSEG.
MRBrainS 2013 Synthetic CircuitFaker .
, .
Diabetic Retinopathy.
Bone Age COPDGene.
GSP
OAI dataset
Kaggle COVID-19
NIR dataset
Hyperspectral Skin Cancer
Dataset
DRIVE
Yunnan & Chongqing datasets Custom (Geant4/GATE MIMIC-CXR RTP CT.
NLST
Kaggle Brain Tumor Dataset UK Biobank Simulated datasets .
, .
, .
BUSI Dataset Coronary.
Fish Corneal.
Rat Retina Real Subjects dataset Public endoscopic images CIFAR-10 ABIDE.
OASIS.
SRPBS
Multi-focus dataset Private (SWIR) Category Public Dataset Public Dataset Public Dataset Private Dataset Public Dataset Synthetic / Simulated Dataset Public Dataset Synthetic / Simulated Dataset Public Dataset Public Dataset Public Dataset Public Dataset Public Dataset Public Dataset Public Dataset Public Dataset Synthetic / Simulated Dataset Public Dataset Public Dataset Public Dataset Public Dataset Synthetic / Simulated Dataset Public Dataset Public Dataset Public Dataset Public Dataset Public Dataset Public Dataset Public Dataset Private Dataset Ref Dataset 27 Rheumatoid Arthritis Patients LIDC/IDRI NCCID CXR.
NCT-CRC-HE-100K
T1 MRI dataset AITEX Fabric Dataset Kvasir-SEG MedMNIST Brain Tumor MRI Dataset EchoNet Mosmed.
Coronacases.
Radiopedia COBRE.
ABIDE
Cervical cells.
RBCs.
ISBI 2014
Lumpy Object Model.
ADNI MRI
Polarimetric Dataset Prostate Cancer Dataset Custom infrared dataset AAPM-Mayo LDCT BraTS'13.
BraTS'18 NIH Chest X-ray 14 Private WLC video K-16.
K-19.
CRC-TP
BrEaST.
BUSI.
Thammasat.
HMSS
Guangzhou CXR dataset NFBS.
CC359.
LPBA.
IBSR
iChallenge-AMD.
ODIR-2019.
RIADD.
STARE
Synthetic & Phantom datasets .
, .
8,092 panoramic radiographs BUS2017.
DDTI.
LiTS.
ATLAS.
BRATS 2015 Shanghai Sixth PeopleAos Hospital Custom & Public datasets IMR & YA-CT images .
Category Public Dataset Public Dataset Public Dataset Public Dataset Public Dataset Public Dataset Public Dataset Public Dataset Public Dataset Public Dataset Public Dataset Public Dataset Public Dataset Synthetic / Simulated Dataset Public Dataset Public Dataset Private Dataset Public Dataset Public Dataset Public Dataset Private Dataset Public Dataset Public Dataset Public Dataset Public Dataset Public Dataset Synthetic / Simulated Dataset Public Dataset Public Dataset Private Dataset Private Dataset Public Dataset Evaluation Method The evaluation methods used in GAN-based biomedical imaging studies reflect a combination of quantitative image quality metrics, task-specific performance measures, and, in some cases, qualitative or expert-based assessments.
Among the most frequently used metrics are PSNR (Peak Signal-toNoise Rati.
and SSIM (Structural Similarity Inde.
, which serve as standard indicators of image reconstruction quality and structural fidelity.
PSNR evaluates pixel-level differences between generated and reference images, making it a useful benchmark for super-resolution, denoising, and image harmonization tasks.
SSIM complements PSNR by assessing perceived image quality based on luminance, contrast, and structural similarity, which is particularly important in medical imaging where structural integrity is critical for diagnosis.
Studies often report both metrics together to provide a more comprehensive evaluation of generative performance.
In diagnostic and classification-oriented applications, evaluation methods frequently include Accuracy.
Precision.
Recall.
Specificity.
F1-score, and AUC (Area Under the ROC Curv.
These metrics are crucial when GANs are used for tasks such as disease detection, tumor classification, or segmentation, as they quantify the balance between detecting true positive cases and avoiding false positives.
For Yuri Pamungkas.
Generative Adversarial Networks in Medical Imaging: A Review of Emerging Trends and Insights Journal of Robotics and Control (JRC) ISSN: 2715-5072 segmentation-specific studies, overlap-based metrics like the Dice Coefficient and Intersection over Union (IoU) are widely used, as they directly measure the degree of correspondence between predicted and ground truth regions.
Some studies also incorporate Hausdorff distance to capture spatial discrepancies between segmentation boundaries, which is important in applications like tumor or organ delineation where boundary accuracy can impact treatment Beyond these common metrics, advanced generative evaluation techniques such as FID (Frychet Inception Distanc.
LPIPS (Learned Perceptual Image Patch Similarit.
, and IS (Inception Scor.
are used to assess the realism and diversity of GAN-generated images.
FID, in particular, is popular in recent studies because it measures the distance between the feature distributions of real and generated images, offering a robust indicator of visual LPIPS focuses on perceptual similarity, which better aligns with human judgment, while IS estimates the diversity and quality of generated samples, albeit with certain limitations in medical domains where class distributions are highly imbalanced.
Some specialized applications integrate domain-specific evaluation methods.
For example, in radiotherapy dose mapping.
Mean Relative Error (MRE) and dose-specific deviation metrics are used.
In elastography studies.
Rago scores and perceptual evaluation by clinicians are included to assess diagnostic relevance.
Similarly, multimodal registration tasks employ registration accuracy measures such as mean absolute error (MAE) in spatial alignment.
rare cases, user studies and expert grading are used alongside quantitative metrics, especially when evaluating subtle visual improvements that are difficult to capture numerically.
Overall, the evaluation strategies reflect the dual need to assess both image realism and task performance.
Image-level metrics like PSNR.
SSIM, and FID ensure that generated images maintain visual and structural fidelity, while taskspecific metrics like Dice.
AUC, and sensitivity confirm that these images are clinically useful.
This multi-metric approach is essential in medical imaging, where visual plausibility alone is insufficient, outputs must also meet rigorous clinical accuracy standards before they can be adopted in real-world diagnostic or therapeutic workflows.
TABLE V.
EVALUATION METRICS MAPPING
Evaluation Metric Primary Purpose F1-score Classification / Detection AUC (Area Under the ROC Curv.
Dice Coefficient Classification / Detection Segmentation Intersection over Union (IoU) Segmentation Hausdorff Distance MAE (Mean Absolute Erro.
MSE (Mean Squared Erro.
RMSE (Root Mean Squared Erro.
nRMSE (Normalized Root Mean Squared Erro.
Rago Score Confusion Matrix Segmentation Regression / Spatial Alignment Image Quality / Regression Image Quality / Regression Description Measures pixel-level similarity where higher values indicate a closer match to reference images Assesses luminance, contrast, and structure similarity which aligns with perceived image quality Measures distribution similarity between generated and real images in feature Estimates perceptual similarity based on deep network features and aligns with human judgment Evaluates both the quality and diversity of generated images Proportion of correctly classified samples over all samples Proportion of predicted positives that are actually positive Proportion of actual positives that are correctly identified Proportion of actual negatives that are correctly identified Harmonic mean of precision and recall to balance false positives and false Measures the ability to distinguish between classes across thresholds Measures the overlap between predicted and ground truth segmentation masks Ratio of intersection area to union area between predicted and reference Measures the maximum distance between segmentation boundaries Average absolute difference between predicted and true values or positions Average squared difference between predicted and reference pixel values Square root of MSE which retains the unit scale of the data Image Quality / Regression RMSE normalized by the range or mean of the data Clinical Assessment Classification / Detection mAP (Mean Average Precisio.
Detection / Segmentation mAP50 / mAP75 Visual Inspection Perceptual Study t-SNE Detection / Segmentation Perceptual Assessment Perceptual Assessment Feature Analysis Mode Score Generative Diversity AM Score BER (Bit Error Rat.
Generative Diversity Signal Analysis MTF (Modulation Transfer Functio.
Image Quality (Optic.
Cos_sim (Cosine Similarit.
Feature Similarity VBM (Voxel-Based Morphometr.
Neuroimaging Analysis Expert grading for thyroid elastography based on stiffness patterns Table showing true positives false positives true negatives and false negatives Average precision over multiple IoU thresholds often used in object detection Average precision at fixed IoU thresholds of 50 percent and 75 percent Expert-based qualitative evaluation of generated images Human observer study to assess visual realism or diagnostic acceptability Visualizes similarity in feature embeddings between generated and real data Measures coverage and quality of generated samples relative to the dataset Evaluates realism and diversity as an adaptation of IS for medical data Measures data accuracy in signal reconstruction tasks Quantifies the system's ability to reproduce contrast at different spatial Measures similarity between feature vectors of generated and reference Statistical analysis of brain anatomy differences between groups PSNR (Peak Signal-to-Noise Rati.
Image Quality SSIM (Structural Similarity Inde.
Image Quality FID (Frychet Inception Distanc.
Image Quality / Perceptual LPIPS (Learned Perceptual Image Patch Similarit.
IS (Inception Scor.
Accuracy Precision Recall (Sensitivit.
Specificity Perceptual Quality Image Quality / Diversity Classification / Detection Classification / Detection Classification / Detection Classification / Detection Yuri Pamungkas.
Generative Adversarial Networks in Medical Imaging: A Review of Emerging Trends and Insights Journal of Robotics and Control (JRC) IV.
ISSN: 2715-5072
STRENGTHS AND LIMITATIONS
Strengths One of the most notable strengths across the studies is the ability of GAN-based models to generate high-fidelity and realistic medical images.
Many works, such as those using StyleGAN.
PGGAN, and StyleGAN3 architectures, demonstrate exceptional visual quality with high SSIM.
PSNR, and low FID scores.
This capability is particularly valuable for tasks like mammogram synthesis .
, retinal fundus image generation .
, and 3D image synthesis .
, where realistic representation is critical for clinical By producing synthetic images that closely mimic real data, these models provide effective solutions to data scarcity and facilitate training of robust diagnostic systems without overfitting.
Another recurring strength is the improvement in downstream diagnostic and segmentation performance through data augmentation and domain adaptation.
Several studies employing cGANs and CycleGANs have demonstrated significant boosts in classification accuracy.
F1-scores, and AUC values when synthetic data are incorporated into the training process.
For instance, lung cancer and pulmonary embolism detection .
and brain tumor segmentation .
benefited from augmented datasets, which enhanced model generalizability to diverse patient populations and imaging conditions.
Additionally, the unpaired image-to-image translation capabilities of CycleGANs allow models to adapt to different imaging domains, such as cross-scanner harmonization in MRI .
or distribution shift correction in CT imaging .
, without the need for paired datasets, a major advantage in real-world medical settings.
Many studies also highlight the capacity of GAN-based approaches to enhance image quality for improved clinical In Medical Physics & Reconstruction tasks, such as low-dose CT denoising .
MRI super-resolution .
, and multimodal image fusion .
GANs consistently outperformed conventional reconstruction methods.
The integration of perceptual loss functions, texture-preserving modules .
TextureWGAN .
), and physics-informed designs .
Physics-assisted GAN .
) has led to sharper, noise-free images that preserve diagnostically relevant This not only aids human interpretation but also improves the performance of automated AI pipelines in detection and segmentation.
A further strength lies in the architectural flexibility of GANs, enabling tailored solutions for specific medical imaging challenges.
Modified architectures like HA-GAN .
for high-resolution volumetric synthesis.
HieGAN .
for multi-scale representation, and SCGAN .
for structure-constrained enhancement address domain-specific For example.
HA-GAN successfully generates 3D CT and MRI volumes with improved structural fidelity, while SCGAN optimizes ophthalmic fundus imaging by integrating quality assessment metrics into the training loop.
These custom adaptations demonstrate the versatility of GAN frameworks in accommodating diverse biomedical imaging Finally, the robustness of GAN models in handling limited or imbalanced datasets stands out as a consistent Semi-supervised approaches like AC-GAN .
and rAC-GAN .
leverage both labeled and unlabeled data, effectively mitigating annotation costs while still delivering competitive performance.
This is particularly beneficial in clinical contexts where expert-annotated data are scarce or expensive to obtain.
Moreover, federated and collaborative GAN learning strategies .
allow various organizations to jointly develop models collaboratively without exchanging unprocessed patient information, ensuring compliance with privacy regulations while expanding data diversity.
Limitations A repeated constraint observed in numerous investigations is the dependency of GAN performance on the caliber and heterogeneity of the training dataset.
While GANs are effective in generating realistic images, their outputs can degrade when trained on small, homogeneous, or biased Several studies, particularly those in niche medical domains such as holographic microscopy .
, fabric defect detection .
, and certain ultrasound applications .
, highlight that limited sample size led to overfitting, reduced generalizability, and artifacts in synthetic images.
In low-data scenarios, models such as DCGAN and early cGAN implementations struggled to maintain structural fidelity, producing outputs that may not be clinically reliable.
Another significant limitation is the risk of generating anatomically inaccurate or clinically misleading features, especially in tasks requiring precise structural preservation, such as tumor segmentation .
, .
or multimodal registration .
Even when visual metrics like SSIM and PSNR are high, subtle anatomical distortions or hallucinated features may occur.
These errors are particularly critical in diagnostic workflows, as they could lead to misinterpretation by radiologists or degrade the performance of AI-assisted diagnosis tools.
For example, texture inconsistencies in TextureWGAN outputs .
and boundary inaccuracies in CycleGAN-based segmentation tasks .
, .
illustrate this Training instability and high computational requirements also appear as common drawbacks.
GAN models, especially high-capacity architectures like StyleGAN2-ADA .
StyleGAN-XL .
, and PGGAN .
, require extensive tuning of hyperparameters, careful balance between generator and discriminator updates, and powerful GPU Several studies reported mode collapse or unstable convergence during training, necessitating the use of advanced regularization techniques such as gradient penalty (WGAN-GP .
) or adaptive discriminator augmentation.
These complexities can hinder reproducibility and limit the accessibility of GAN-based methods in clinical research environments with constrained computational resources.
A further limitation lies in the lack of standardized evaluation protocols across studies.
While metrics like SSIM.
PSNR, and FID are widely used, they primarily assess visual similarity rather than clinical validity.
Only a subset of works, such as those in ophthalmology .
, .
and cardiac imaging .
, incorporated domain-specific or expert-based The absence of clinically grounded benchmarks makes it difficult to compare results across studies and raises concerns about the real-world diagnostic impact of GAN- Yuri Pamungkas.
Generative Adversarial Networks in Medical Imaging: A Review of Emerging Trends and Insights Journal of Robotics and Control (JRC) ISSN: 2715-5072 generated outputs.
Finally, regulatory and ethical challenges remain underexplored in most papers.
Although some studies address privacy concerns through federated learning .
or data anonymization, very few consider the downstream implications of deploying GAN-generated images in clinical decision-making.
The potential for misuse, such as generating realistic but synthetic patient images that could be mistaken for real data, presents ethical risks that require formal guidelines and verification processes before clinical CONCLUSION In conclusion, the systematic review of GAN-based approaches in biomedical imaging demonstrates that these architectures .
anging from foundational models like DCGAN and cGAN to advanced variants such as StyleGAN.
PGGAN.
WGAN-GP, and specialized modified GAN.
have significantly advanced capabilities in image synthesis, enhancement, and cross-domain adaptation.
Their applications span a broad spectrum of biomedical domains, with the highest concentration in Medical Physics & Reconstruction and Cancer Imaging & Diagnosis, supported by datasets both public and proprietary, including CBISDDSM.
ChestX-ray14.
BraTS.
OAI, and ABIDE.
Evaluation commonly relies on quantitative image quality metrics such as SSIM.
PSNR, and FID, supplemented by task-specific measures like Dice.
AUC, and expert visual assessment.
The key strengths of these methods lie in their ability to generate realistic high-quality images, improve diagnostic performance through data augmentation, enhance imaging in low-quality or low-dose conditions, and adapt flexibly to diverse clinical tasks, even with limited labeled data.
Nonetheless, limitations persist, including dependency on high-quality and diverse datasets, risks of anatomical inaccuracy, training instability, high computational demands, non-standardized evaluation protocols, and underexplored ethical and regulatory considerations.
VI.
FUTURE DIRECTIONS
Future research on GANs in biomedical imaging should focus on bridging the gap between technical advancements and clinical translation.
Priority should be given to developing architectures that not only generate visually realistic outputs but also preserve fine-grained anatomical accuracy, ensuring clinical safety and trustworthiness.
Integrating domain knowledge, such as physics-informed constraints and pathology-specific priors, could further enhance the interpretability and reliability of generated Standardized, clinically relevant evaluation protocols, combining quantitative metrics with expert-driven assessments, are needed to allow fair benchmarking across studies and modalities.
Additionally, expanding research into underexplored domains such as histopathology, vascular imaging, infectious diseases, and pediatric imaging could unlock new diagnostic opportunities.
Addressing computational efficiency and training stability will be crucial for enabling deployment in resource-limited healthcare settings, while privacy-preserving approaches like federated GAN learning should be advanced to support multiinstitutional collaboration without compromising patient Finally, ethical and regulatory frameworks must evolve in parallel with technical innovation to mitigate risks of misuse and to define clear guidelines for the safe integration of GAN-generated content into clinical ACKNOWLEDGMENT The authors would like to acknowledge the Department of Medical Technology.
Institut Teknologi Sepuluh Nopember, for the facilities and support in this research.
The authors also gratefully acknowledge financial support from the Institut Teknologi Sepuluh Nopember for this work, under project scheme of the Publication Writing and IPR Incentive Program (PPHKI) 2025.
REFERENCES