Recent advancements in computing power and big data analysis have significantly boosted the development of Artificial Intelligence (AI)—systems that mimic human cognitive abilities, such as learning, problem-solving, and decision-making [18–22]. AI can process and analyze large datasets efficiently. Machine Learning (ML), a subset of AI, learns from data by identifying patterns and features [23]. Two fundamental types of machine learning are supervised learning and unsupervised learning. While supervised learning [24,25] requires labeled data for training, unsupervised learning discovers patterns in unlabeled data, making it more applicable in scenarios where labeling is infeasible. Among these, supervised learning is the most widely utilized and successful approach. In supervised learning, algorithms are provided with a data set comprising pairs of input and output examples. The algorithm learns to map each input with its corresponding output, effectively associating input examples to output examples. A widely used form of supervised learning is classification. Once trained, supervised learning algorithms can achieve accuracy levels that exceed human performance, making them essential in various products and services. Despite these advancements, the learning process has limitations compared to human abilities. Current supervised learning approaches typically require millions of training examples [26]. To address these challenges, researchers are increasingly focusing on unsupervised learning, to reduce dependence on extensive human supervision and decrease the number of training examples needed. In general, the purpose of unsupervised learning is to extract meaningful information from a data set containing unlabeled input examples. Unlike supervised learning, unsupervised learning seeks to uncover useful patterns from unlabeled data [27,28]. Two well-known applications of unsupervised learning are clustering and dimensionality reduction.

A significant approach in unsupervised learning is generative modeling. Generative modeling aims to approximate the true data distribution pdata(x) with a model distribution pmodel(x). This is achieved by designing a function pmodel(x;y) with parameters y and optimizing these parameters to make pmodel(x) as similar to pdata(x) as possible. A common technique for generative modeling is maximum likelihood estimation, which minimizes the Kullback-Leibler (KL) divergence between pdata(x) and pmodel(x). Traditionally, explicit density models with simple probabilistic forms were used for such tasks. However, with advances in machine learning, more complex models have been developed to handle high-dimensional data. Diffusion models have recently gained attention in the field of computer vision for their impressive performance in generative tasks. These models consist of a two-phase process: a forward noise injection phase and a reverse reconstruction phase. During the forward phase, the original data is corrupted by adding Gaussian noise. The reverse phase consists of a neural network trained to reconstruct the original data by denoising it. Although these models can produce high quality and diverse outputs, they are computationally intensive and often suffer from slow generation times [29]. To address these limitations, Goodfellow et al. [30] introduced Generative Adversarial Networks (GANs) [31], a novel generative model that has become a prominent tool for tasks like image denoising, translation, segmentation, and reconstruction. Generative Adversarial Networks (GANs) [32–34] enables unsupervised learning by generating new data samples from an existing data distribution.

Formally, a GAN consists of two models: a generator G(z;Wg) and a discriminator D(x;Wd). The generator maps noise vectors z∼pz(z) to the data space, creating synthetic samples G(z). The discriminator outputs a probability D(x)∈[0,1] representing the likelihood that (x) came from the real data distribution pdata(x) rather than G(z). The two models are trained simultaneously in a minimax game and the equation as follows [30]: (1)minGmaxDV(D,G)=Ex∼pdata(x)[log⁡D(x)]+Ez∼pz(z)[log⁡(1−D(G(z)))]

This adversarial process drives the generator to improve its outputs such that they are indistinguishable from real samples, while the discriminator becomes better at detection.

In general, the two neural networks: a generator and a discriminator, are designed to compete with each other [35]. Fig. 3 illustrates the structure of a GAN. The generator and discriminator architectures generally consist of multi-layer convolutional or fully connected layers. The generator learns the statistical properties of the training data and generates new images, while the discriminator evaluates and distinguishes between real and synthetic images [30]. Both networks serve as mappings between data domains [36]. The generator, without direct access to the real dataset, aims to create convincing synthetic images to deceive the discriminator. If the discriminator makes an incorrect classification, an error signal is generated to refine the generator’s output, progressively enhancing the quality of generated images. The generator transforms a latent space into the data space, while the discriminator maps image data to a probability score, indicating whether an image is real or synthetic. If the discriminator identifies an image as real, it outputs a value close to 1, whereas for a synthetic image, it outputs a value near 0. Fig. 4 shows the training process of a GAN.

Over the years, GANs have undergone substantial development, starting from the original adversarial learning concept, although initial versions struggled with fully capturing the data distribution [42]. A comparison of some popular GAN models is mentioned in Table 1. The original GAN, commonly referred to as the Vanilla GAN [37], utilizes random noise as input to the generator, which synthesizes photorealistic images. The discriminator differentiates between real and generated (fake) images. In the absence of ground truth images, the generator is trained solely using adversarial loss, while the discriminator is optimized using classification loss. However, the Vanilla GAN is limited in its ability to generate images across diverse classes effectively. Training Vanilla GANs is inherently unstable and often results in generators yielding outputs that lack coherence structure. A set of architectural constraints was proposed to overcome the un-stability issue and evaluated for Convolutional GANs, termed Deep Convolutional GANs (DCGAN) [38]. Trained discriminators demonstrate competitive performance in image classification tasks compared to other unsupervised methods. Visualization of learned filters reveals that specific filters specialize in generating distinct objects. Compared to Vanilla GAN, DCGAN replaces pooling layers with strided convolutions in the discriminator and fractional-strided convolutions in the generator. Additionally, it incorporates batch normalization in both the generator and discriminator. The emergence of DCGAN brought improvements in image fidelity [38], and WGAN later addressed challenges related to mode collapse and training instability [40]. Vanilla GAN and DCGAN exhibit several limitations, one of which is the inability to control the generated outputs. For example, while a GAN can train a generator to produce images of digits (0−9) from random noise, practical applications often require generating a specific image. This limitation can be addressed by incorporating an additional input to guide the generation process. Previously, the generative model was pg(x). Now, it is designed to produce pg(x|c), where c is a conditional input used to control the generation process. This input c can be a string of codes representing the desired output or intent. GAN models often encounter challenges like mode collapse, which hinder their ability to provide meaningful learning curves that are crucial for debugging and hyper-parameter tuning. This issue can be addressed by using Wasserstein GAN (WGAN) [40], which incorporates the Wasserstein distance, also known as the Earth Mover’s distance. In the WGAN architecture, the discriminator is replaced with a critic, which evaluates the Wasserstein distance rather than performing binary classification.

GANs generate images by sampling from a probability distribution, while CycleGANs [41] perform image-to-image translation between two domains. CycleGAN establishes a mapping G:X→Y such that the output y^=G(x),x∈X, is indistinguishable from real images y∈Y by an adversary trained to differentiate y^ from y. Unlike conventional GANs, CycleGAN employs two generators and two discriminators to enable bidirectional translation between the domains. To enhance translation quality, CycleGAN incorporates cycle consistency loss in addition to adversarial loss, ensuring that a translated image can be mapped back to its original domain. CycleGAN enabled transformations between image domains without the need for paired datasets [41], and PGAN introduced a stepwise training method to generate increasingly detailed images [43]. Subsequent models like SAGAN concentrated on identifying important image areas and capturing long-range dependencies [41]. More recent innovations include RANDGAN, which improves segmentation for anomaly detection and outperforms earlier GAN frameworks in the medical context [44]. DGGAN focuses on generating anonymized brain vascular imagery using MRA patches [45], and ED−GAN integrates VAEs with GANs to enhance generative performance [46]. These advancements collectively demonstrate the substantial progress of GANs, broadening their applicability in areas such as medical image synthesis, anomaly detection, and complex data representation.

With the variety of GAN models available for different applications, the base model or its specific variants can be selected based on the task. For instance, DCGAN is suitable for image generation, SRGAN for super-resolution, U-Net-based GANs for segmentation, and CycleGAN for image translation.

GANs [47–50] have demonstrated their potential in various fields, including image generation, style transfer, and data augmentation. Their ability to generate high-fidelity synthetic images without extensive labeled datasets makes them valuable for medical imaging. The following are the applications of GANs [51] in medical imaging:

Image Denoising: Medical images, such as low-dose CT scans or accelerated MRI, often have lower quality due to noise or reduced resolution. GANs enhance these images by reducing noise and increasing resolution [39,52–54].

The use of GANs in medical imaging represents a significant improvement in healthcare. GANs offer distinct advantages in addressing the challenges of medical imaging, primarily due to their capacity to produce highly realistic images, even with limited supervision which is in contrast to conventional models that focus solely on classification or detection tasks. This is important in the medical domain, where labeled datasets are often scarce, imbalanced, or costly to acquire. In addition, GANs are effective in tasks that require one form of medical image to be transformed into another, such as improving the resolution of scans, converting between imaging modalities. These applications are vital for accurate clinical assessment.

Ongoing research and technological progress continue to expand the scope and impact of GANs in healthcare [82]. For instance, CycleGAN is employed for domain adaptation, facilitating the translation of images across different modalities [41], while pix2pix supports tasks like resolution improvement and denoising through image-to-image translation [83]. A further example is UNITGAN, which allows cross-modal image integration by establishing a shared latent space between modalities. In contrast, ProGAN contributes to the generation of high-resolution images [43]. These GAN models have played a key role in enhancing both the diversity and quality of medical images, which in turn supports improved healthcare analysis and outcomes. Fig. 5 summarizes the application of GAN in medical imaging.

Despite their potential, GANs present certain challenges. Training GANs can be computationally demanding, and achieving stable training requires careful model design and parameter tuning. The interpretability of GAN-based models is another concern, particularly in high-stakes medical applications where model transparency is essential. Ethical considerations are also crucial, especially regarding patient privacy and the potential misuse of synthetic medical images.

This research adopts a Systematic Literature Review (SLR) approach, guided by the methodologies introduced by Kitchenham et al. [84]. The study design follows the structured process outlined in the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) framework. This ensures a clear, consistent, and reproducible process for gathering and analyzing data. All key literature included in this review was selected in accordance with PRISMA's widely accepted standards for conducting systematic reviews. The following steps were included in the search strategy. 1.

Databases Searched: The review targets scholarly, peer-reviewed articles sourced from well-established databases, including ScienceDirect, SpringerLink, the ACM Digital Library, IEEE Xplore, PubMed, Scopus and Web of Science. At the outset of this study, a total of 500 articles were initially selected for review. The articles were sourced from literature published between 2020 and 2025, with the final selection process carried out between November 2024 and January 2025. Following a thorough evaluation, 167 articles were shortlisted based on their relevance to applications, challenges, and recent advancements. The deliberate focus on recently published literature reflects our commitment to providing a forward-looking and state-of-the-art analysis.

This approach ensured a comprehensive, reproducible, and methodologically sound search process in alignment with PRISMA standards. The PRISMA flow diagram of the article selection procedure is shown in Fig. 6. PRISMA checklists can be found in the supplementary files.

The remaining sections of this paper explore in detail how GANs are applied within the field of medical imaging. Section 2 discusses the various applications of GANs in medical images. Section 4 details the different challenges of GANs for clinical use under regulatory frameworks. Section 5 gives an insight to the other genrative models present. Finally, the Section 6 concludes the paper providing an insight into the limitations and future of GANs in the medical field.

Image denoising is a vital preprocessing act in analysis, as all types of medical images are susceptible to noise [100–104]. The sources of noise in medical imaging can be categorized into sensor-related, acquisition-related, and radiation-related factors. Computed Tomography (CT) is a widely utilized technique for disease diagnosis, but it carries the potential risk of radiation exposure [105,106]. Reducing radiation levels can lead to increased noise in CT images. Reconstruction of Low-Dose CT (LDCT) images offers an effective approach to address this issue. Fig. 7 shows a GAN based framework for medical image denoising. A GAN utilizing Wasserstein distance (WGAN) and perceptual similarity was applied to denoise CT images in [94]. The perceptual loss minimized noise by aligning output and ground truth features in a defined space, while the GAN shifted the noise distribution from strong to weak. Wasserstein distance was used to compare the distributions of normal-dose CT (NDCT) and LDCT data. Feature extraction was performed using Convolutional Neural Networks (CNN) based on the Visual Geometry Group (VGG). Performance metrics such as PSNR and SSIM were employed to evaluate the outputs of different networks, with WGAN−VGG achieving superior performance. Huang et al. in [107] presented a denoising method (DU−GAN) which utilized U-Net-based discriminators to assess the global and local variations in the denoised and normal images. A denoising method based on Conditional GAN was popularized by Li et al. in [53] in which the image context relationship and structural information was preserved. The method was tested on LIDC dataset and was seen to outperform the state of the art works. Zhu et al. had introduced a denoising method based on GAN in [108]. Molecular activity in human tissues was captured using Single-Photon Emission Computed Tomography (SPECT), which relies on gamma rays for image acquisition in [91]. High-noise SPECT images are input into the generator, while the discriminator assesses the generated images against real samples, specifically low-noise SPECT images. The loss, which quantifies the difference between generated and real images, is used to update both the generator and discriminator simultaneously. Both components were optimized using the Adam optimizer, with a learning rate of 0.00001 and 800 training epochs. Noise levels were assessed using the normalized standard deviation (NSD), enabling a comparison of results with and without conditional GAN denoising applied to the reconstructed SPECT images.

Segmentation plays a crucial role in medical image analysis. Automating the segmentation process is highly challenging due to variations in anatomical structures across different patients [55,109–111]. Skin cancer, common among individuals with fair skin, is classified into melanoma (pigmented lesions) and non-melanoma (non-pigmented lesions). Early detection of melanoma is critical. Dermoscopic images captured via smartphones can be analyzed to detect pigmented lesions. A GAN was trained using 3000 color images in [99], with the generator employing a U−Net architecture and the discriminator using the Adam optimizer. Training was conducted in two phases: without image rotation and with image rotation. The highest segmentation accuracy was achieved when training included image rotation. However, the proposed method lacked pre-processing stages. Future improvements in the same work were involved in enhancing the GAN architecture, utilizing larger datasets, and analyzing specific color channels beyond conventional RGB spaces. X-ray imaging, which uses high-frequency electromagnetic waves, depends on the radiological density of tissues to determine the level of absorption. Chest X-rays can detect infections, tuberculosis, cancer, and other chronic chest conditions. Segmenting postero-anterior chest X-ray images involves isolating the left and right lung fields and the heart. Since chest X-rays are 3D projections onto 2D images, overlapping structures complicate segmentation. The Structure Correcting Adversarial Network (SCAN) framework [99] applied adversarial techniques for semantic segmentation, comprising a segmentation network and a critic network that were trained jointly. Both the segmentation and critic networks in SCAN employed Fully Convolutional Networks (FCNs). While FCNs were effective for RGB images, SCAN used FCNs for grayscale chest X-ray images, enabling extraction of high-level data representations. Metrics such as Intersection-over-Union (IoU) and Dice coefficient were used to evaluate performance. The SCAN framework was trained on 247 chest X-ray images from the Japanese Society of Radiological Technology (JSRT) dataset, which included 154 images with lung nodules and 93 without. Additionally, 138 chest X-ray images from the Montgomery dataset were used, with 117 for training and 21 for testing. Approximate IoU values for the JSRT and Montgomery datasets were 95.1% and 93%, respectively. This method represented a successful application of convolutional neural networks for accurate segmentation results. Spinal MR image enabled the detection of defects in the spinal cord, vertebrae, intervertebral discs, and more. The Spine−GAN [57] method employed a dynamic optimization algorithm combined with a hybrid learning strategy. The Spine−GAN included a segmentation network with three components: an encoder, a local long short-term memory module (LSTM), and a decoder. An Atrous Convolutional Autoencoder (ACAE), was designed for spinal image representation and pixel-level classification. The LSTM, a recurrent neural network (RNN), modelled spinal structural details by leveraging spatial correlations, thereby enhancing network performance. A Convolutional Neural Network (CNN) discriminator mitigated overfitting and incorporated a robust learning strategy with a flexible optimization algorithm. The discriminator accepted inputs from either the Segmentor or the ground truth. Spinal MR data from 253 patients, captured using 1.5T equipment, was used for training and testing, comprising 5343 samples of neural foramen images, 1818 disc images, and vertebrae images. Data was split into five subgroups, with one group used for training and the others for testing. Performance metrics included pixel-level accuracy, Dice coefficient, specificity, and sensitivity, yielding an accuracy of 96.2%, specificity of 89.1%, and sensitivity of 86%. Limitations include its inapplicability to all MR image types and the lack of prior clinical spine knowledge in the diagnosis. Future improvements could address these limitations.

Fig. 8 shows a framework for medical image segmentation using GAN. A healthy immune system defends against foreign bodies in the body. In autoimmune disorders, the immune system attacks healthy tissues. Human epithelial type 2 (HEp−2) cell images can be analyzed to diagnose autoimmune disorders [95]. Due to the large variety of patterns, the available datasets were limited. cC−GAN, a transfer learning framework in GAN, is a powerful segmentation approach for HEp−2 datasets, addressing overfitting and improving transfer capacity. The cC−GAN used three losses: L1 Loss for label prediction, GAN Loss to help the discriminator distinguish between outputs, and Softmax Loss for classification. The generator in cC−GAN employed Residual U−Net (RU−Net) architecture, while an additional epoch training scheme (AEt) was used for stable training of the generator and discriminator. Two datasets were used: the first dataset contained 252 specimens from six categories (Homogeneous, Speckled, Nucleolar, Centromere, Golgi, Nuclear membrane, and Mitotic spindle), while the second dataset contained 28 green-channel HEp−2 images with six categories (Centromere, Homogeneous, Fine Speckled, Coarse Speckled, Nucleolar, and Cytoplasmic). The smaller size of the second dataset impacted fine-tuning, making it less effective for segmentation. Segmentation accuracy and precision were used for evaluation, with cC−GAN achieving 86.15% accuracy compared to Fully Convolutional ResNet (FCRN), which achieved 87.29%. The FCRN’s deeper network architecture improved accuracy but also increased the risk of overfitting due to a higher number of parameters. cC−GAN effectively distinguished between real and synthetic samples and classified cells. A multi-scale L1 loss function was effective for semantic segmentation tasks [56]. Training for SegAN followed a conventional GAN min-max framework, involving a Segmentor network (S) and a Critic network (C). Both networks were trained using back propagation, optimizing the multi-scale L1 loss. The Segmentor is a fully convolutional encoder-decoder with a kernel size of 4×4 and a stride of 2 for downsampling. The same loss function was applied for training both the Segmentor and the Critic. The method was evaluated using 220 high-grade and 54 low-grade brain MR images, with Dice, Precision, and Sensitivity serving as evaluation metrics. SegAN compared the segmented image with the ground truth across multiple critic layers. Xue et al. in [112] had introduced a method for multi organ thorax segmentation using U−GAN.

The neural network architecture proposed in [113] is composed of two main components: a Segmentor and a Critic. To enhance semantic feature extraction, a novel model named Transformer−CV−Unet(TCUnet) is introduced. TCUnet was utilized as the generator within a GAN framework to perform image segmentation, improving both robustness and computational efficiency. Once the generator segmented the target images, the Critic integrated the latent features with hierarchical information from various modalities. Additionally, a hybrid adversarial mechanism incorporating a multi-phase CV energy functional is developed. This integrated framework, referred to as AdvTCUnet, leverages the strengths of both the generator and critic modules. Comprehensive evaluations on the BraTs 2019−2021 datasets demonstrate that the proposed method outperforms current leading approaches in brain tumor segmentation on MRI scans, achieving Dice Similarity Coefficients of 0.8642 for Enhancing Tumor (ET), 0.9303 for Whole Tumor (WT), and 0.9060 for Tumor Core (TC) on BraTs2021. A Local Cross-Attention Unet (LCAUnet) was introduced by Wang et al. in [114] for the segmentation of skin lesions using edge and body features.

High-resolution images play a crucial role in accurate disease diagnosis. Despite advancements in medical imaging techniques, factors such as imaging conditions, equipment limitations, and environmental obstructions can result in low-resolution images [115–119]. Resolution can be improved by various methods such as enhancing the spatial resolution, interpolation, multi image super resolution methods. MR image resolution can be improved using GAN technique by taking advantage of the volumetric information in the 3DMR image [59,120]. The proposed architecture to improve the MRI resolution was based on Super Resolution GAN (SRGAN). The difference between conventional GAN and SRGAN is in the convolutional filters, they are 3D convolutional layers capable to handle volumetric information. To overcome the problem of vanishing gradient in cross entropy loss function, adversarial loss was used in the discriminator and in generator network, combination of adversarial and content loss is used. Generator network contains had six residual blocks with 32 convolutional filters of size 3×3×3. The discriminator network had eight convolutional layers of kernel size 3×3×3. 589T1−weighted images (470 for training and 119 for testing) from the Alzheimers Disease Neuro imaging Initiative (ADNI) database was used and skull stripping was done for all images. Adam optimization was employed in both generator and discriminator. The generated super resolution images for upsampling factor of 2,4 were compared against the ground truth in terms of of peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM). Both metrics yielded a better value for the upsampling factor of 2.

Histopathology is the study of presence of disease in biopsy or surgical samples using microscopes [121]. Histopathological images provides a comprehensive view of cell tissues. Different segments in a tissue are visualized by pigmenting it using different dyes. Histopathological images analysis involves segmentation, detection, feature extraction, classification, etc. A unified GAN framework with a newly designed loss function was introduced for the same in [96]. The loss function is derived from both WGAN with gradient penalty (WGAN−GP) and Information Maximizing Generative Adversarial Networks (InfoGAN). During training, the generator (G) aimed to estimate the distribution Pg such that it aligns with the real data distribution Pr. This estimation was done by optimizing the Earth Mover (EM) distance approximation between generator (G) and discriminator (D). In the meantime the auxiliary network (Q) maximized the mutual information between the chosen random variables and the samples generated by Q. In the testing phase Q was capable of classifying as softmax function as the final layer in the Q network. Proposed unified architecture was modelled as a pipeline comprising of the following four stages-Nuclei segmentation, Clustering/classification of cell by Q network, calculating the cell proportions and image level prediction based on cell proportions, generating interpretable cell categories by generator for different noises. Four different bone marrow datasets were examined for testing and training the network. Images from these datasets were further sub divided into cell-level images and were classified into categories such as neutrophils, myeloblasts, monocytes and lymphocytes. While training, 32 or 64 Gaussian noise variables were used and 64 categorical variable were used. While segmenting the cell images, to separate out the touched cells, morphological opening with kernel size 7×7 was carried out. Feature extraction process could be disturbed by the color and feature contrast, this could be avoided by using non segmented histopathological images. By using such images, cells could overlap drastically. The problem of overlapping cells could be excluded by using a bounding box of size 32×32 on non segmented image. For each cell class, precision, recall and F-Score were the segmentation evaluation parameters. Cell clustering evaluation metrics were purity, entropy and F-score. High purity value with low entropy indicated better clustering. Gradient penalty required to compute the second order derivative, so is really time consuming while training. The proposed architecture could be improvised by changing the segmentation method. It could be further improved by considering patient related data such as clinical trials, gene expression data. Anatomical segmentation of lesions and locating pathology becomes complex when anatomy or pathology is small like retinal images, cardiac MR, etc. or the image quality is too low due the acquisition process [92].

Diabetic retinopathy, glaucoma are the diseases which are diagnosed with the help of retinal fundus images. Retinal fundus image resolution are not good enough to detect microaneurysms, hemorrhages as they cover small image areas. Progressive GAN (P−GAN) generated a high resolution image from a low resolution input image. The proposed P−GAN architecture incorporated two stages-stage 1 and stage 2, with output from stage 1 being the input to the stage 2. Triplet loss function ensured quality improvement in images as they progressed from one stage to another. Low resolution version from high resolution image could be acquired by the application of Gaussian filter followed by down sampling. The proposed network could yield better quality images for high scaling factors, greater than 8. The proposed architecture with 2 stages have a set of generator (G1) and discriminator (D1) in stage 1 and another set of generator (G2) and discriminator (D2) in stage 2. The loss function for the stage 1 include Mean square error (MSE) and CNN loss terms. In contrast to stage 1, the loss function for stage 2 is the triplet loss function. Every generator block the network contains a up sampler with factor 2. Triplet loss function is a minmax function such that it minimizes the distance between output image from stage 2 and ground truth and maximizes the distance between output image from stage 2 and output image from stage 1. This dual constrains made sure that improvement was achieved both qualitatively and quantitatively. The network wass trained with 5000 retinal fundus images and tested with another 1000 images. The dataset was down sampled by factors 2, 4, 8, 16, 32 and for each sampling the time taken to generate super resolved image was computed. As the up sampling factor increases, there was evident increase in the time to generate super resolved images. The proposed P−GAN was compared with various SR−GAN algorithms. SSIM, PSNR (dB), sharpness metric (S) [122] were the metrics for network performance evaluation and was computed by varying the scaling factor and for various noise types like Gaussian, salt and pepper noise and speckle noise. With the increase in upsampling factor, a decrease in SSIM, PSNR, S3 metrics was estimated. Out of the three noises, low quality images with Gaussian noise, only yielded the expected results. Vessel segmentation with P−GAN outperform all other GAN algorithms and was compared with respect to segmentation accuracy, Specificity, area under curve (AUC) and sensitivity. The proposed P−GAN had been evaluated using cardiac MR dataset for improving the resolution and for cardiac LV segmentation.

X-Ray Computed Tomography (X-Ray CT) uses X-Rays for screening, diagnosis, image guided surgery, etc. The quality of X-Ray CT images can be improved in two ways-either by using a good quality hardware or by computationally enhancing the images. The former is not economical and also are of high radiations [123]. High radiations cause gene damages and can cause cancer. Low Dose CT (LDCT) uses lower radiation, which results in low quality CT images. Computational techniques can be employed to obtain High Resolution CT (HRCT) from LDCT. A novel residual CNN-based network in the CycleGAN framework known as Deep Cycle-Consistent Adversarial SRCT network (GAN−CIRCLE) [93] was one among the computational techniques which preserved the anatomical details in CT images. The proposed algorithm had the following advantages over the other conventional GAN architectures

cycle consistency to ensures strong across domain consistency between LRCT and HRCT,

The main challenges in recovering HRCT images from noisy LRCT images are-complex spatial variations in the images, presence of unique noise patterns, sampling and degradation makes the image blur. To handle these limitations, non linear SR functional blocks with residual module is included in the proposed framework, which have the ability to learn high frequency details. Adversarial learning in a cyclic manner is followed, which ensures superior quality CT images. The proposed network has two generators—G & F, with feature extraction network and reconstruction network. Reconstruction network has a Parallelized CNN with multi layer perceptron (MLP) to carry out nonlinear projection in the spatial domain. Parallelized CNN are network within network which can perform dimensionality reduction with faster computation and less information loss, able to learn the complex mapping at finer levels with better accuracy. The framework has been trained and tested using two datasets-twenty five images from Tibia dataset and 5936 images from Abdominal dataset. Performance of proposed GAN−CIRCLE is compared with various GAN using the metrics PSNR, SSIM and Information fidelity criterion (IFC). Enhancing the image quality acquired by portable miniature devices is of great attention recently.

A deep learning approach named MedSRGAN was proposed to enhance the resolution of medical images using Generative Adversarial Networks (GANs). Fig. 9 shows the architecture of MedSRGAN. At the core of this method is a specially designed generator known as the Residual Whole Map Attention Network (RWMAN), which effectively extracts important features across multiple channels while emphasizing critical regions within the images. To optimize performance, a composite loss function was utilized—combining content loss, adversarial loss, and feature-based adversarial loss. The MedSRGAN model was trained and evaluated using a dataset comprising 242 thoracic CT scans and 110 brain MR scans. Experimental results indicated that the method not only maintains fine texture details but also produces more lifelike and accurate high-resolution images. Furthermore, a Mean Opinion Score (MOS) assessment conducted by five experienced radiologists on CT image slices confirmed the model’s effectiveness.

Degraded images are the major drawback of portable imaging devices. Spatial resolution, contrast and noise are the three issues related with ultrasound images. Portable devices produce low resolution, low contrast and high noise images, which makes the disease diagnosis inaccurate. GAN based method provides promising results for resolution enhancement with the following advantages-non linear multi level mapping between LR and HR images, adaptive feature extraction without human intervention, image quality enhancement using discriminator D, direct and efficient one step feed forward reconstruction procedure and easier implementation on hardware like FPGAs. The encoder-decoder based GAN has the problem of bottleneck which constrain the information sharing between LR image and HR image. To eliminate this problem, U-net based architecture is employed successfully. U-net model lacks in performance due to the presence of speckle noise in the HR image. In order to overcome the problems of encoder decoder model and U-net model, sparse skip connection UNet (SSCU−Net) is proposed [125]. Local patches in the low resolution images are used in the discriminator. Assuming these patches are independent, discriminator is capable of modeling high frequency details. The input image patches makes the training easier and lower the memory constraints. The proposed network uses two losses-L1 loss for conserving the low frequency information and differential loss for high frequency information like edge sharpness. Dataset for testing and training the proposed network is done using 50 simulation pairs, 72 vivo image pairs and 40 phantom pair. Each dataset is divided into five groups—four groups for training and one group for testing. PSNR, SSIM, contrast resolution (CR) and mutual information (MI) are estimated for the performance evaluation. PSNR, SSIM measure similarity between LR and HR images, higher the PSNR, higher intensity similarity. CR estimates the ability to differentiate the intensity difference. Even though the SSIM for U-net is higher, the U-net images have a over smoothed appearance which implies the loss of some high frequency details. The proposed SSCU-Net out performed the U-Net and the encoder-decoder model and also the images have better resolution with preserving more edge details. Luan et al. in [126] have introduced a deep learning based adaptive matching network (AM-Net) alongwith a dataset generation method named Multi-mapping (MMP) for Ultrasound Localization Microscopy (ULM). A GAN based MR image super resolution technique was proposed in [127] that uses a generator mechanism incorporating multiple feature selection methods. Jia et al. in [128] have introduced a super resolution method for retinal fundus image based on GANs which uses a vascular structure prior. This is shown to have overcome the shortcomings of its previous model Real−ESRGAN [129].

Disease diagnosis and treatment with single image modality may not be adequate in most of the cases. The images acquired cannot outline the complete anatomical details or fails to acquire the details with the desired imaging modality [85,90,130–132]. Image translation is an optimum solution, where required image is synthesized from a different image modality, without inducing much cost or risk. The challenge involved in translating one image modality to another is the presence of unrealistic data in the output image [133]. Fully supervised learning methods are among the most widely used deep learning approaches for this task [134]. However, these methods require paired low- and high-quality images for training, which is especially challenging in medical imaging, where obtaining such aligned image pairs is difficult in real-world scenarios. To address this limitation, several unsupervised learning frameworks have been developed [132,135]. Despite their potential, these approaches often face issues such as instability, noise amplification, and the occurrence of halo artifacts. A well-known solution for unpaired image-to-image translation is the Cycle-consistent Generative Adversarial Network (CycleGAN) [100,136]. The block diagram of (CycleGAN) for image-to-image translation between domains A and B is shown in Fig. 10. This architecture enables the model to learn domain-specific knowledge from representative images and transfer it to another domain without requiring paired training images. Nevertheless, most bidirectional GAN-based models are insufficiently constrained. For instance, while CycleGAN excels at capturing inter-domain cycle-consistency and global appearance within domains, it often struggles to preserve local details. This limitation is particularly significant in medical imaging, where precise local details are essential for accurate decision-making. High-quality medical images are expected to have uniform illumination and well-defined structural details to support effective diagnosis.

A skin lesion synthesizer based on GAN was proposed in [98], having a coarse to fine generator, multi scale discriminator and a robust objective function for learning. This method synthesized images from a semantic label map and an instance map [60]. Images of resolution 1024×512 were synthesized with proposed generator which incorporated convolutional layers, few residual blocks and deconvolutional layers. The multi scale discriminator consisted of three discriminators, with same input of different resolutions. The training was stabilized using the feature matching loss function, that compared the real and synthetic image features from all the discriminators. The loss function for the proposed network included conditional GAN loss and feature matching loss. Clinically meaningful 2594 skin lesion images were split into two-2346 training images and 248 testing images. Quantitative analysis was performed using AUC and p-value. Multiple images of same anatomy with different contrast improves disease diagnosis easier. Multi contrast MR image synthesizer with conditional GAN(cGAN) was proposed in [61] for spatially registered images. cGAN was compared with cyclic GAN for registered and unregistered T1, T2MRI images. T1 weighted MRI image described the gray and white matters, T2 images described cortical tissue fluid. The method based on pGAN was also evaluated in the same. pGAN consists of a generator, pre-trained V GG network and a discriminator. The generator learned to synthesize T2-weighted image from the T1-weighted image. Simultaneously the discriminator differentiated between the real and synthetic image. Pixel by pixel, adversarial and perceptual losses were minimized by the generator and discriminator maximized the adversarial loss. pGAN was trained with pixel and perceptual loss. Likewise T1-weighted image was synthesized from T2-weighted image. cGAN implementation used cycle loss instead of pixel wise loss in the pGAN. cGAN used two generators-GT1 and GT2 and two discriminators-DT1 and DT2; one pair of generator and discriminator for T1 image and other for T2 image. GT1 synthesized a T1 weighted image from the respective T2 image, meanwhile DT1 discriminated between real T1 image and synthetic T1 image. Three datasets were used-healthy images from MIDAS, IXI and abnormal images from BRATS. Out of the 66 subjects analyzed from MIDAS dataset, 48 were used for training, 5 for validation and 13 for testing. From each subject, 75 axial cross sections of brain tissues without artifacts were selected manually. IXI dataset used 40 subjects with 25 subjects for training, 5 for validation and 10 for training. 41 low grade giloma patients from BRATS dataset were evaluated where 24 for training, 2 for validation and 15 for testing. To avoid biases on dataset, 40 subjects with 4000−5000 images from each dataset were analyzed. Dataset normalization ensured the absence of bias in quantitative evaluation. Armanious et al. in [90] presented MedGAN which is a framework designed for end to end medical image to image translation at the image level. It leverages recent progress in Generative Adversarial Networks (GANs) by integrating the adversarial approach with a novel blend of non-adversarial loss functions. A key component of the system is a discriminator network that functions as a trainable feature extractor, enforcing alignment between the generated medical images and their target modalities. To ensure accurate reproduction of textures and detailed structures, style transfer losses are incorporated. Furthermore, the architecture introduces a new generator design called CasNet, which progressively refines medical image outputs using a series of encoder-decoder modules to improve image clarity and detail. Also, five experienced radiologists evaluated and verified MedGAN’s outputs quality based on subjective assessment.

Yang et al. in [9] introduced a method to perform Image Modality Translation (IMT) using a deep learning approach grounded in Conditional Generative Adversarial Networks (cGANs). This framework captured low-level pixel information and high-level semantic features such as brain tumors and anatomical structures across different imaging modalities showcasing strong potential as a supportive tool in medical diagnostics. A cross-modality registration technique that integrated deformation fields was introduced, enabling the model to incorporate information from the translated imaging modalities. A Translated Multichannel Segmentation (TMS) approach for MR data. In this method, both original and translated modalities were processed together using Fully Convolutional Networks (FCNs) to perform segmentation in a multichannel fashion. These two approaches effectively leverage cross-modality information to enhance performance without the need for additional data. Fig. 11 illustrates the above method.

A novel GAN architecture named MMTrans, which leverages a Swin Transformer backbone for multi-modal medical image translation, is introduced in [137]. The proposed system is composed of three primary components: a generator, a registration module, and a discriminator. The registration module employs a Swin Transformer to estimate a deformable vector field (DVF) based on the SwinIR framework, which aligns the generated image with the target [138]. For paired datasets, spatial inconsistencies between the input and target images are corrected using this registration network. In scenarios involving unpaired data, the generator G(x) produces outputs that preserve the anatomical structure of T1 images and adopt the visual style of T2. The registered output R(G(x),y) further the generated image to match both the morphology and appearance of T2. A convolutional neural network-based discriminator evaluates whether the generated images are indistinguishable from actual target modality images. Through extensive testing on both publicly available and clinical datasets—covering both paired and unpaired cases—MMTrans demonstrated superior performance compared to existing methods and showed strong potential for clinical adoption. Nonetheless, there are some limitations to this approach. Although processing 2D slices offers computational efficiency, the spatial context provided by 3D medical volumes is crucial for many diagnostic tasks. Consequently, future work should focus on extending MMTrans to operate effectively on 3D data to fully leverage its clinical applicability.

Chen et al. in [139] introduced MI−GAN, a novel multi-domain medical image translation algorithm that incorporates a key transfer branch. By analyzing the imbalance present in medical imaging datasets, the approach identified critical target domain images and constructed a specialized transfer branch. Utilizing a single generator, the method facilitates multi-domain image translation in the medical context. This structure enhanced both the model’s attention mechanism and the quality of the generated images. Additionally, a lung image classification model was presented, leveraging synthetic image data for augmentation. The training dataset combines both synthetic lung CT scans and original real-world images to evaluate the effectiveness of the model in diagnosing normal individuals, as well as patients with mild and severe cases of COVID-19. The method was seen to out perform the state of art techniques. Ozbey et al. have introduced a method in which adversarial diffusion modelling is used for obtaining improved results in image translation [140].

Medical image reconstruction is a crucial process for generating high-quality images needed for accurate analysis. However, the quality of these images is often affected by noise and artifacts [141–143]. To address these limitations, there has been a paradigm shift from traditional analytical and iterative reconstruction methods to data-driven machine learning approaches [144,145].

A review of the literature reveals that frameworks like pix2pix and CycleGAN are widely employed for MR image reconstruction [120]. In this method 3D neural network architecture called the multi-level densely connected super-resolution network (mDCSRN), which incorporates training guidance from a GAN is proposed. The proposed mDCSRN is designed for efficient training and inference, while the GAN component enhances the realism of the super-resolved images, making them nearly indistinguishable from the original high-resolution counterparts. An illustrative representation of this method is given in Fig. 12. Liao et al. [146] explored sparse-view CBCT reconstruction to reduce artifacts, proposing a feature pyramid network for the discriminator and computing a modulated focus map to preserve anatomical structures during reconstruction. Alongside reconstruction from undersampled data, maintaining domain data accuracy is essential. In MR reconstruction, undersampled k-space data in the frequency domain has also been addressed [147,148]. Various types of loss functions have been applied in image reconstruction to capture local image structures effectively. For example, cycle-consistency and identity loss have been utilized together for denoising cardiac CT [149]. Wolterink et al. proposed a method for low-dose CT denoising by excluding some domain loss, but this approach resulted in compromised local image structure [106]. Bhadra et al. in [150] introduced an image-adaptive GAN based reconstruction approach (IAGAN), designed to enhance data fidelity by adjusting the pretrained generative model parameters based on the acquired measurement data. The IAGAN framework is applied to reconstruct images from undersampled MR data. A cutting-edge generative adversarial model, Progressive Growing of GANs(ProGAN), was trained using a large dataset of high-quality images from the NYUfastMRI repository. The trained generator was then integrated into the IAGAN architecture to reconstruct high-resolution images from retrospectively undersampled k−space data in the validation set. The results demonstrate that this GAN driven reconstruction method can recover intricate anatomical details from noisy or incomplete measurements, offering a level of detail that conventional reconstruction techniques—typically dependent on sparsity-based regularization—may struggle to achieve. This highlights the potential of IAGAN in improving the diagnostic value of MR scans.

Additionally, conditional GANs(cGAN) with skip connections in the generator have been used to synthesize full-dose equivalent PET scans from low-dose data was explored by in Rashid et al. in [151]. The primary aim of this study was to assess the effectiveness of a cGAN-based approach in enhancing image quality, minimizing noise, and accelerating reconstruction time, in comparison to conventional methods such as Maximum Likelihood Expectation maximization (MLEM) and Total Variation (TV). The approach involved iterative training of a U-Net based generator with a full-image discriminator. Results demonstrated that the proposed cGAN framework significantly improved image sharpness, reduced noise, and offered faster reconstruction, outperforming traditional techniques.

Ahn et al. in [152] utilized 10,000 anteroposterior (AP) knee radiographs to develop a more cost-effective and balanced medical imaging dataset. Two convolutional neural network models were implemented: Deep Convolutional GAN(DCGAN) and StyleGAN2 with Adaptive Discriminator Augmentation (StyleGAN2−ADA). To assess the images generated by StyleGAN2−ADA compared to actual radiographs, a Visual Turing Test was conducted involving two computer vision specialists, two orthopedic surgeons, and a musculoskeletal radiologist. For evaluation, the Fréchet Inception Distance (FID) and Principal Component Analysis (PCA) were employed. The synthetic images successfully replicated key pathological features such as osteophyte formation, joint space narrowing, and subchondral sclerosis. Expert classification accuracy when distinguishing real from generated images varied, with scores of 34%, 43%, 44%, 57%, and 50%. The FID score between the generated and authentic images was 2.96, significantly lower than that of another medical dataset (BreCaHAD=15.1), indicating higher image fidelity. PCA results revealed no statistically significant differences between principal components of the real and generated images (p>0.05). Overall, this research highlights the potential of GANs in generating realistic images.

A novel framework named DualMMP−GAN was introduced in [153] for generating high fidelity medical images from a given source modality. This method enhances the traditional CycleGAN by incorporating dilated residual blocks, a dual-scale patch based discriminator, and a perceptual consistency loss to improve generation quality, particularly in regions containing lesions. DualMMP−GAN excels in preserving both contextual information and fine structural details, making it more effective in reconstructing lesion areas. Instead of relying on the standard single-scale discriminator used in CycleGAN, DualMMP−GAN adopts a dual-scale discriminator that operates on image patches. This design allows the network to learn features at both fine and coarse levels. The synergy between these two scales improves the model’s adaptability to lesions of varying sizes, shapes, and locations. To enhance the network’s ability to capture contextual features without significantly increasing computational cost, dilated residual blocks are introduced in place of standard residual blocks. These dilated blocks widen the receptive field, enabling extraction of spatial and structural information from MR data. This contributes to more accurate preservation of lesion boundaries, morphology, and overall image continuity, resulting in more detailed and high-resolution outputs. Furthermore, the model leverages perceptual consistency loss, an improvement over the traditional cycle consistency loss employed by CycleGAN. By comparing feature representations across different layers, this loss function ensures that the generated images closely resemble the target modality at multiple levels of abstraction, thereby enhancing visual clarity and detail in the synthesized MR scans.

Synthetic pterygium images were produced using the default configuration of the StyleGAN3 architecture in [154], chosen for its strong performance and advanced generative capabilities, which are well-suited for producing photorealistic outputs. The generator employed a 512-dimensional latent and intermediate space, with a mapping network consisting of two layers specifically adjusted for image synthesis. To maximize the model’s representational power, the channel base was set to 32, 768, capped at 512 channels. An Exponential Moving Average (EMA) value of 0.99 was applied to control parameter updates and promote training stability.

The discriminator also followed the StyleGAN3 design without freezing any layers. A minibatch standard deviation group size of four was used to encourage variation across samples during training. It mirrored the generator in terms of channel base and maximum. Both models were optimized using the Adam optimizer, configured with beta parameters [0,0.99] and an epsilon of 1e−08. Learning rates were assigned as 0.0025 for the generator and 0.002 for the discriminator to maintain stable learning. Loss was utilized alongside an R1 regularization term with a weight of 8.0 to ensure controlled gradient updates within the discriminator.

Despite the architectural strengths of StyleGAN3, generating highly realistic medical images presented certain limitations. Initially, a lack of diversity in the training set contributed to elevated Fréchet Inception Distance (FID) scores, reflecting a degree of synthetic bias. Additionally, the complexity and variability found in medical imagery required meticulous tuning of model parameters and implementation of sophisticated augmentation techniques to improve visual authenticity. Evaluation through confusion matrices demonstrated that the synthetic images reached a high level of realism, though clinician performance varied—highlighting the subjective nature of visual interpretation and the difficulty in achieving universally indistinguishable synthetic outputs.

These outcomes suggest that cGAN based reconstruction holds strong potential to enhance diagnostic precision and streamline clinical imaging workflows. Although GANs have been widely applied in various medical imaging tasks, their direct integration into clinical diagnostics and decision-making remains challenging. A significant portion of current research in image reconstruction relies on conventional quantitative metrics to evaluate performance. However, when GANs are trained using additional loss functions, optimizing for visual quality becomes difficult, particularly in the absence of a specialized reference metric tailored to assess perceptual image quality. This presents a key obstacle to aligning GAN generated outputs with clinical standards [65].

Understanding how diseases evolve over time is essential for early detection and effective treatment planning. This is particularly important for severe conditions like Idiopathic Pulmonary Fibrosis (IPF), a chronic and progressive lung disease with a survival rate similar to that of various cancers. CT scans are widely recognized as a dependable method for diagnosing IPF. Predicting future CT images for patients in the early stages of IPF can play a significant role in enhancing treatment strategies and improving patient outcomes. Zhao et al. introduced a novel model in [155] named 4D Vector Quantised Generative Adversarial Networks (4D−VQ−GAN) designed to synthesize realistic CT volumes of IPF patients across different time points. The model training process involved two main stages. Initially, a 3D−VQ−GAN was trained to reconstruct CT volumes. Subsequently, a temporal model based on Neural Ordinary Differential Equations (ODEs) was trained to learn the temporal progression of the quantized latent embeddings produced by the first stage’s encoder. Various configurations of the models were tested to generate sequential CT scans and compare to actual data using both quantitative metrics and qualitative assessments. Survival analysis based on imaging biomarkers extracted from the synthetic CT volumes were performed to validate the clinical relevance of our generated scans. The resulting concordance index (C−index) was on par with that of biomarkers obtained from real CT scans, indicating the potential of generated CT data for accurate survival prediction and real-world clinical application.

Deep learning models typically require large datasets, making it difficult to apply them in situations where limited data is available. One common solution is data augmentation, which involves creating training examples by generating new data. This approach involves basic techniques like random rotations, flipping, cropping, and adding noise. However, such transformations often fall short when applied to complex datasets like medical images. To address this, researchers have developed more strategies for medical imaging. The primary aim is to produce synthetic data that closely mirrors the original distribution. The emergence of Generative Adversarial Networks (GANs) has significantly enhanced data augmentation capabilities [70].

GAN based data augmentation involves training a generator network to produce synthetic images from a latent space, thereby increasing the dataset’s diversity and variability beyond simple transformations. This method is especially advantageous in situations with limited labeled data or class imbalance, where creating additional samples of underrepresented classes can greatly enhance classifier performance [156]. For instance, in medical image analysis, GANs have been employed to generate rare pathological cases, facilitating better diagnostic model training without the need for extensive manual data collection [148].

A data augmentation technique was presented by Frid et al. in [157] data augmentation approach that integrates traditional image perturbation techniques with the generation of synthetic liver lesions using Generative Adversarial Networks (GANs) to enhance liver lesion classification accuracy. The main contributions of this work include: (1) the generation of high-quality synthetic focal liver lesions from CT scans using GANs, (2) the development of a convolutional neural network (CNN)-based model for liver lesion classification, and (3) the enrichment of the CNN training dataset with synthetic samples to achieve improved classification performance. The work compares the results obtained with classical augmentation (not involving the usage og GAN). It could be seen that the classification performance improved progressively with the increase in training data, reaching a plateau at approximately 78.6%, beyond which the inclusion of additional augmented samples did not lead to further enhancement in accuracy.

Gan et al. introduced a generative adversarial network (GAN) model, named HieGAN in [158], which employs a hierarchical structure to produce high-quality synthetic knee images. This model is intended to enhance data augmentation strategies for deep learning tasks. During the training phase, HieGAN incorporates an attention mechanism within both the generator and discriminator networks, specifically before the 256×256 image scale, to better extract critical features from knee images. To ensure stable training, a novel approach combining pixelwise and spectral normalization was applied. HieGAN was assessed using a large-scale dataset of knee images, with performance measured by AmScore and ModeScore. Experimental results demonstrated that HieGAN surpassed existing state-of-the-art methods. Consequently, HieGAN represents a promising advancement toward developing more reliable deep learning models for knee image segmentation. Future research should explore clinical validation through Visual Turing Tests.

Furthermore, datasets augmented with GANs have demonstrated improved performance of deep learning models by reducing overfitting and increasing robustness to noise and variations in data. The adversarial training process pushes the generator to create samples that challenge the discriminator, resulting in realistic synthetic data that enriches the training set [159]. However, challenges persist in ensuring the quality and diversity of GAN generated samples, as mode collapse and training instability can hinder their effectiveness.

In GANs, achieving convergence between the generator and discriminator at a global optimum—known as the Nash equilibrium—is essential. The training process follows a minimax game framework aimed at reaching this equilibrium. Effective training strategies for both networks are crucial for optimal performance. As the generator becomes more proficient, it generates synthetic images that closely resemble real ones, making it increasingly difficult for the discriminator to tell them apart. When the generator reaches peak performance and produces highly realistic images, the discriminator’s accuracy drops to around 50%, indicating it can no longer differentiate between real and fake data. At this stage, the feedback provided to the generator becomes uninformative, hindering further improvement in image quality. This imbalance can ultimately result in non-convergence during GAN training [166]. The issue of non-convergence significantly impacts the quality of synthetic image generation. This problem becomes evident when examining the characteristics of the generated outputs. In many cases, non-convergence causes the generator to fail, resulting in the production of flat, single-color images—such as entirely black or white—particularly in gray scale image synthesis.

The issue of non-convergence is a significant challenge encountered during the training of GANs. To address the persistent issue of non-convergence in biomedical imaging applications of GANs, various technical studies have been examined. One proposed solution involves guiding the training process toward achieving a Nash equilibrium, which can help stabilize GAN training. However, maintaining this equilibrium during training is notably challenging. With this concept as a foundation, the surveyed literature is organized into three principal categories [166]:

optimization of update algorithms [167]

Mode collapse represents a significant challenge during the training of GANs. This issue often results in outputs that lack the diversity observed in real medical images. When mode collapse occurs, the generator overlooks important features and generates identical patterns, leading to a loss of meaningful variability in the synthetic data. This repetition reduces the utility of the generated images. Training a GAN to completely eliminate mode collapse remains a difficult task. For instance, when generating segmented chest radiographs using ground truth data and corresponding segmentation masks, mode collapse can lead to repetitive or incomplete outputs. The high complexity and dimensionality of 3D brain MR data further contribute to the occurrence of mode collapse during image synthesis, making it a persistent challenge in this domain.

GANs may also introduce feature hallucinations while generating synthetic data [174]. Feature hallucination refers to the creation of artificial elements or the omission of critical features in generated images, which can lead to diagnostic errors [175]. These hallucinations often arise as a result of mode collapse during training. This issue is prominent in image to image translation tasks, where synthetic outputs may include inaccurate or misleading features. Addressing mode collapse effectively can help minimize the occurrence of hallucinated features in GAN generated images. Various strategies have been employed in GANs to tackle the issue of mode collapse. Ensuring training of both the generator and discriminator is necessary. The generator must be able to capture the complete range of feature distributions and anatomical structures present, while the discriminator must provide feedback so that the generator can produce diverse outputs. Exploring advanced optimization methods to improve the stability of the training process by resolving challenges like mode collapse can significantly boost the performance of GANs.

The mode collapse problem can be alleviated by using different methods such as

regularization

In addressing the training challenges of GANs, evaluation metrics play a vital role in assessing the diversity and quality of the generated images. To measure image diversity, particularly in the context of mode collapse, metrics such as Inception Score (IS), Multi-Scale Structural Similarity Index (MS−SSIM), and Maximum Mean Discrepancy (MMD) are commonly used. For issues like non-convergence and instability, Peak Signal-to-Noise Ratio (PSNR) and Fréchet Inception Distance (FID) are frequently applied. Among these, IS and FID are standard metrics for evaluating image quality; however, both are based on models pretrained on the ImageNet dataset [191], which does not include biomedical image classes. As a result, these metrics are not ideally suited for applications in the biomedical imaging field. Additionally, MS−SSIM is a perceptual metric focused primarily on luminance and contrast, offering a limited view of image similarity. PSNR, while widely adopted for assessing image quality, is primarily effective for grayscale images and lacks robustness in more complex scenarios. In the field of biomedical image analysis, researchers often rely on conventional pixel-wise evaluation metrics to assess the performance of GANs. These metrics are typically designed for supervised learning scenarios that depend on the presence of reference images. However, in biomedical imaging, acquiring reference data is challenging due to privacy concerns and the errors associated with manual annotations. Hence, unsupervised learning methods are commonly used. Also, assessing the performance of GANs is crucial due to factors such as random initialization, optimization variability, and technical complexities. Comparing generated images to real ones remains a difficult task, highlighting the need for further investigation into reliable evaluation techniques. Consequently, developing evaluation metrics that capture both subjective perceptions and objective measures presents a significant challenge for the field.

Medical imaging concerns handling highly sensitive patient data, which raises important privacy concerns during both data collection and application. While GANs are capable of generating visually convincing medical images, a major issue is the introduction of artificial elements known as hallucinated features. These artifacts may mimic real pathological signs, leading to incorrect diagnostic algorithms. These can compromise clinical reliability. Therefore, it is essential to validate synthetic outputs through both expert evaluation and robust quantitative testing to ensure they reflect true features. To improve the integrity of these images, it is essential to integrate physics-informed simulations and conduct thorough experimental evaluations aimed at understanding the convergence behavior of GANs within the medical imaging domain. Using GAN generated data in clinical settings requires adherence to strict regulatory standards. Health authorities such as the Food and Drug Administration (FDA), USA and European Medicines Agency (EMA) demand clear evidence of performance, safety, and reliability to show that the synthetic images are clinically valid, reproducible, and free from bias. Documentation of the development process, ethical oversight, and compliance with data protection laws (e.g., HIPAA, GDPR) are also crucial. Additionally, synthetic images must be clearly labeled, and their use justified with comprehensive risk assessments to ensure transparency.

Federated learning presents a promising approach to reduce privacy concerns while using GANs in medical imaging. However, it encounters several challenges. Achieving effective model convergence under the constraints of restricted communication bandwidth in decentralized environments remains a major difficulty. Also, combining federated learning with differential privacy requires a careful balance between safeguarding sensitive data and maintaining high model performance. The complexity of medical data because of its heterogeneity and the variety of imaging modalities complicates the collaborative model training. Future research should lay priority on overcoming communication inefficiencies, improving the resilience of models, and customizing federated learning techniques to accommodate the characteristics of medical datasets. Furthermore, exploring advanced privacy preserving methods and striving to create a fully decentralized, privacy centric GAN capable of generating high-quality medical images without compromising patient confidentiality continues to be a challenging yet vital endeavor in the healthcare sector.

Fig. 13 gives an illustration of the steps in federated learning. In the first phase, each participant independently calculates the model gradients on their local data. To ensure data confidentiality, cryptographic methods such as homomorphic encryption are applied before transmitting the encrypted gradients to the central server. In the second phase, the central (master) server performs secure aggregation of the encrypted gradients. In the third phase, the aggregated results are sent back to the participants. During the fourth phase, each participant decrypts the aggregated gradients and updates their local model accordingly. This cycle is repeated iteratively until either the loss function reaches convergence or a predefined number of iterations is completed. Throughout this process, the participants’ data remains stored locally, maintaining privacy and offering an advantage over centralized approaches such as those based on Hadoop. Federated learning facilitates collaborative model training across multiple databases without the need to centralize the data, enabling scalability with growing datasets while minimizing communication overhead, as only gradients—not raw data—are shared.

GAN based image generation models can produce photorealistic images; however, in medical applications, caution is essential as these images play a critical role in disease detection and diagnosis. It is imperative for medical professionals, as the end users, to thoroughly evaluate and validate GAN generated images to ensure their reliability and clinical utility. Trust and acceptance from doctors are crucial for integrating such models into healthcare workflows. While PSNR and SSIM are standard for evaluating image processing techniques [192], they often fail to reflect perceptual quality or clinical relevance, particularly in the presence of subtle distortions. This underscores the importance of expert-driven evaluation, as automated metrics alone may not capture diagnostic integrity. In medical imaging, radiologists and clinicians are best positioned to perform this qualitative assessment, as their expertise ensures that reconstructed images are not only visually plausible but also diagnostically accurate.

Realistic full-field digital mammograms were generated using a progressive GAN architecture [193], achieving high resolution and realism indistinguishable from real images, even by domain experts. Despite the specialized nature of medical imaging, both experts and non-experts in a reader study showed random success probabilities, emphasizing the critical role of human validation in ensuring clinical applicability. Similarly, the progressive growing GAN (PGGAN) [194] was employed to generate high-resolution chest radiographs (CXRs), with two models trained separately on normal and abnormal CXRs, producing synthetic images of 1000×1000 pixels. Six radiologists performed binary Turing tests on two validation sets, attaining mean accuracies of 67.42% and 69.92% in the first and second trials, respectively. Majority voting and Cohen’s score were used to further assess diagnostic agreement and reliability. Another notable contribution is MedGAN [90], an end-to-end image translation framework tailored for medical applications. By integrating a conditional adversarial setup with multiple non-adversarial losses and a CasNET generator, MedGAN enhances global consistency and preserves high-frequency details. Its effectiveness was confirmed through expert radiologist evaluations, indicating strong clinical fidelity.

Further advancing GAN applications, GANCS [195] presents a compressed sensing framework that models the low-dimensional manifold of high-quality MR images using a least-squares GAN (LSGAN) to capture fine textures, combined with ℓ1/ℓ2 loss to suppress high-frequency noise. Expert radiologist ratings on a large paediatric contrast-enhanced MR dataset consistently preferred GANCS over traditional wavelet, dictionary learning, and pixel-wise deep learning methods. Diagnostic reliability was supported by normalized Radiologist Opinion Scores (ROS) for image quality, artifact presence, and sharpness, aligning with SSIM and SNR metrics. Additionally, a hybrid deep learning reconstruction strategy was proposed by integrating CycleGAN with projection onto convex sets (POCS) [196]. The initial CycleGAN output was refined using POCS and reused in a second training iteration to enhance performance. Radiologists evaluated five reconstruction methods (U−Net, GAN, CycleGAN, RefineGAN, and POCS−CycleGAN) across brain and knee MRI datasets. POCS−CycleGAN achieved the highest average scores for both knee (3.3 and 2.7) and brain MR images (3.7 and 4.05), consistently demonstrating superior image quality across different modalities and readers.