Content area
Medical imaging is transformed with Generative artificial intelligence (AI) that offers robust tools for image synthesis, data augmentation and enhancement of image quality. Generative Adversarial Networks (GANs) have established themselves among the various generative models as particularly robust in synthesizing realistic medical images near real-world clinical data. This review discusses the growing importance of generative AI in synthesizing medical images, its use in applications like radiology, pathology and other medical disciplines. We present the overview of some of the significant generative models, e.g., Variational Autoencoders (VAEs) and Diffusion Models and their advantages, disadvantages and prospects. We also present the challenges that accompany such models like interpretability, transferability to other medical disciplines and the ethics of applying synthetic data to real-world clinical practice. Further, the review presents recent developments in hybrid AI approaches that combine AI and physics-based models along with multimodal learning in an attempt to enhance the trustworthiness and accuracy of generative methods. Finally, we look forward to future research directions like federated learning and explainable AI (XAI) that will enable the safe and successful application of generative AI in medicine. XAI methods—such as visual attribution (e.g., Grad-CAM, SHAP), latent space interpretability in VAEs, and external symbolic explanation frameworks—are particularly important in improving trust and understanding in clinical applications [Huff et al., 2021; Bhati et al., 2024]. The aim of this paper is to facilitate researchers and practitioners to showcase the full potential of generative AI for medical imaging by presenting a comprehensive review of existing techniques and emerging trends.
Introduction
Medical imaging is essential in modern healthcare for diagnosis, therapy planning, and monitoring disease progression, with MRI, CT, and X-rays widely used in fields like radiology, pathology, and oncology [1]. These modalities enable non-invasive visualization but face challenges like high acquisition costs, lack of annotated data, and image variability due to noise or motion [2]. AI, especially generative models, is helping address these issues by generating realistic synthetic data, enhancing image quality, and enabling cross-modality translation (e.g., MRI to CT) [3, 5, 12]. This AI-medical imaging convergence is transforming clinical practice and reducing disparities [6]. Generative AI began with Variational Autoencoders (VAEs), which introduced probabilistic latent representations for image synthesis [8]. A major leap occurred with GANs, proposed by Goodfellow et al. in 2014, using generator–discriminator networks to produce high-fidelity images [8, 10]. GANs were soon adopted in healthcare, such as by Nie et al. for generating CT images from MRI data [48]. Over time, GAN architectures evolved to produce higher-resolution images [3]. Recently, Diffusion Models have emerged, excelling in denoising and generating 3D images, as shown by Khader et al. [26]. These advances mark a shift from manual image processing to deep learning–driven synthesis, offering powerful solutions to persistent imaging challenges [13, 14].
[See PDF for image]
Fig. 1
Distribution of research across medical imaging modalities (2015–2023), derived from a systematic search of PubMed, Scopus, Web of Science, and IEEE Xplore [18, 78, 79]
Figure 1 shows research statistics on various imaging modalities and Fig. 2 shows published research on generative AI (2015–2023). This review highlights the pivotal role of generative AI in medical image synthesis, a rapidly advancing field with strong clinical relevance [15]. GANs have revolutionized the generation of high-fidelity medical images, as demonstrated by Beers et al. [3]. Complementary models like VAEs and Diffusion Models offer strengths such as better interpretability and noise robustness [16]. Chen et al. further emphasized the realism achievable with Diffusion Models [5]. This review aims to unify these advancements, focusing on real-world applications and limitations [17].
[See PDF for image]
Fig. 2
Growth in research areas (2015–2023) for generative AI models, based on a systematic search of PubMed, Scopus, Web of Science, and IEEE Xplore [18, 80, 81]
As AI becomes integral to healthcare, generative models are key for addressing data gaps, improving image quality, and enabling new diagnostic approaches [18]. Studies like Han et al.’s survey of GAN-based synthesis serve as vital resources for researchers and practitioners [18, 20]. This review critically evaluates how generative AI enhances medical imaging across radiology, pathology, and oncology [4, 21, 32]. It provides a comprehensive overview of key models—GANs, VAEs, and Diffusion Models—focusing on their technical foundations, applications, and limitations [22]. Emphasis is placed on their role in addressing data scarcity and image quality issues, from early developments to advanced use cases like brain tumor classification using hybrid VAE-GANs [2, 23]. The paper is structured into five sections: introduction, model overview, clinical need, technical progress, and application-specific insights with future directions [24]. It also addresses challenges in interpretability, domain adaptation, and ethics, as noted by Koohi-Moghadam and Bae [30], offering a clear roadmap for understanding the role of generative AI in healthcare [26]. Generative AI is reshaping healthcare by addressing key challenges in medical imaging and beyond [27]. A major application is data augmentation, where synthetic images help improve diagnostic model training, as shown by Madani et al. in chest X-ray classification [39]. It also aids in preserving patient privacy by generating synthetic data to reduce reliance on real records [30]. With growing demands for precision and speed, generative AI enhances image quality (e.g., super-resolution by Sun and Liao) and supports cross-modality translation like MRI-to-CT synthesis. This is crucial for rare diseases and low-resource settings where data is limited [33]. Beyond imaging, it also plays a role in drug discovery, treatment prediction, and clinical trial simulation, marking it as a future pillar of healthcare innovation [20, 34]. The future of generative AI in clinical imaging looks promising, driven by advancements like federated learning and explainable AI [35]. Federated learning enables privacy-preserving, decentralized model training, ideal for synthetic data generation and collaborative research under strict data regulations [36, 37]. Explainable AI improves model interpretability, addressing a key barrier to clinical adoption [30]. Other innovations include physics-based constraint hybrids (e.g., Kim and Lee’s synthetic MRI) and multimodal systems combining imaging with other data types (e.g., Chowdhery and Narang) [8, 28]. These developments aim to enhance the reliability and generalizability of generative AI, paving the way for safer clinical use and advancing precision medicine [40, 41].
Generative AI in medical image synthesis
Generative artificial intelligence (AI) has become a revolutionary power in medical image synthesis which is providing cutting-edge tools that generate synthetic images to optimize clinical workflows, enhance limited datasets and improve diagnostic accuracy. This area keeps advancing rapidly with the power of deep learning and rising needs for high-quality medical imaging. In this section, the primary generative models i.e. Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs) and Diffusion Models are described in detail, discussing their mechanisms, strengths, weaknesses and medical imaging applications. Their technical background and recent directions are investigated along with their revolutionary potential in healthcare imaging and a complete background for understanding their application in contemporary medicine. When we refer to explainability in the context of generative models for medical imaging, it includes visual attribution techniques (such as Grad-CAM, SHAP), interpretability of latent spaces in VAEs, and adversarial approaches that highlight critical regions influencing image generation [Huff et al., 2021; Bhati et al., 2024; Qin et al., 2024]. These are essential for making AI outputs clinically reliable.
Methodologies for generative AI in medical image synthesis
Generative AI is transforming medical imaging by providing innovative tools that not only replicate the quality of real medical scans but also fill gaps in clinical data, streamline workflows, and enhance diagnostic reliability. At the heart of this technology are three foundational model families—Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), and Diffusion Models—each with unique mechanisms and strengths that address different challenges in medical image synthesis [3].
GANs function through a dynamic “game” between two neural networks: the generator, which creates synthetic images, and the discriminator, which evaluates their realism. This adversarial setup helps the generator progressively improve until the output becomes nearly indistinguishable from genuine medical scans. For example, progressively grown GANs have been successfully used to produce high-resolution images that retain intricate anatomical features, a breakthrough for diagnostic imaging where visual precision is crucial [3, 7]. These models have even been trained on contextual mammogram data, showing their capability to replicate the subtle patterns necessary for breast cancer detection [19, 59]. Further innovations, such as style-based GANs, provide fine control over image features like contrast and texture, enabling the creation of tailored synthetic images that meet the specific needs of different diagnostic applications, including brain imaging [18, 25]. VAEs, on the other hand, take a probabilistic approach. They encode input images into a compressed latent space and decode them back, allowing the generation of new, coherent images. This structure makes them highly effective for unsupervised learning and scalable modeling—particularly useful in domains with limited labeled data. For instance, conditional VAEs have been used to detect pathologies in medical images without requiring extensive annotations, making them valuable in studying rare diseases or emerging conditions [20]. While VAEs may not always preserve fine details as well as GANs, their ability to model data distributions and capture broader structural features remains a strong asset [36]. Diffusion Models represent a newer and highly promising direction. These models simulate a gradual denoising process: starting from a noisy version of an image, they iteratively learn to reconstruct clean, high-fidelity outputs. This method has proven particularly effective in tasks like cross-modality translation, such as generating CT-like images from MRI scans—offering practical benefits like reduced radiation exposure and fewer required imaging sessions [29, 42]. Enhancements to these models, such as focusing on critical spatial frequencies during the denoising process, have further improved the sharpness and diagnostic relevance of the images they produce [27]. Some newer diffusion approaches even incorporate physics-informed CNNs, aligning the synthetic output with real-world imaging principles like magnetic resonance physics, adding an extra layer of reliability [28].
Discussion on classification of generative AI models
From Fig. 3 it has been found that Image generative AI, based on density modeling approaches, can broadly be classified into two categories: implicit density models and explicit density models. Implicit density models do not explicitly model probability distributions and include Generative Adversarial Networks (GANs) and diffusion models such as LDM and DDIM. In contrast, explicit density models explicitly model the data distribution and are divided into tractable and non-tractable models. Tractable models, which are computationally feasible, include auto-regressive models like PixelCNNs and flow-based models that use invertible transformations. Non-tractable models, which involve approximate or complex likelihood estimation, include Variational Auto-Encoders (VAEs), energy-based models, and early diffusion models such as DDPM. The subsequent sections will focus on application-specific use cases and clinical integration, minimizing redundancy of model mechanics already discussed here.
[See PDF for image]
Fig. 3
Classification of generative AI models on the basis of density
Applications of generative AI in medical image synthesis
Generative AI models such as GANs, VAEs, and Diffusion Models are revolutionizing medical imaging through applications like data augmentation, cross-modality synthesis, and image enhancement.
[See PDF for image]
Fig. 4
Image synthesis using different synthetic samples to improve generalization
Figure 4 shows the image synthesis process using different synthetic samples to improve generalization of image synthesis. GANs generate photorealistic images ideal for tumor visualization in radiology, with improved stability using switchable CycleGANs for CT kernel conversion [17], though artifacts may occur if miscalibrated [5]. VAEs offer interpretable latent spaces for synthetic image verification, as shown in endoscopic image synthesis [11], though limited by lower image resolution [13]. Diffusion Models enhance image quality through iterative denoising, useful in low-dose CT scans [9, 31, 34], despite high computational costs [26]. Applications include MRI data synthesis for segmentation via PSIGAN [23], high-resolution radiology images from progressive GANs [3], and MRI-to-CT translation using semi-supervised GANs [24]. Diffusion Models improve low-quality scans [9], and GANs aid COVID-19 detection in X-rays [43]. RadImageGAN combines MRI, CT, and patient data for improved synthesis [37], while physics-based constraints enhance clinical reliability [28]. Multimodal learning integrates imaging with genomic data for refined outputs [10]. Federated learning supports privacy-preserving synthesis [44], and frequency-guided Diffusion Models enable zero-shot translation without extensive data [12]. Explainable AI increases model transparency, aiding clinical adoption [30], while generative AI improves diagnostic accuracy and expands applications across medical domains [30]. Real-world applications further highlight the practical relevance of these technologies. The RODEO autoencoder, developed by Mehta and Majumdar [45], delivers real-time image reconstruction, a critical capability in fast-paced clinical environments like emergency rooms. In another study, a multi-level noise-aware GAN architecture was applied to enhance low-dose CT images, helping preserve diagnostic quality while reducing patient radiation exposure [46]. Moon et al. [47] emphasized the importance of data diversity in training sets—especially in diseases like glioma—showing that AI models trained on a wide range of phenotypes achieve better and more equitable diagnostic outcomes. These models are helps to speed up imaging processes without sacrificing quality. Okolie et al. [49] used generative models to reduce breast MRI acquisition time, while Onishi and Tanaka [50] applied GANs to improve pulmonary nodule detection in CT scans—offering earlier and more accurate lung cancer diagnoses. Even research from adjacent fields, such as DeepID-Net in computer vision, has influenced medical AI by advancing object detection capabilities critical to recognizing subtle features in complex medical images [51].
The role of generative AI in medical image synthesis
The use of generative AI in medical image synthesis is no longer just an innovation—it’s becoming a necessity. Healthcare professionals and researchers often face significant hurdles: access to diverse and high-quality imaging data can be severely limited, especially in specialized areas like pediatric care, oncology, or rare diseases such as cystic fibrosis. Traditional imaging datasets are frequently incomplete, imbalanced, or too small to train robust diagnostic models. In such cases, generative models offer a powerful alternative. Technologies like Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs), and Diffusion Models are reshaping the landscape of medical imaging. These tools don’t just simulate data; they generate clinically meaningful, high-fidelity images that support complex diagnostics. For instance, during the early stages of the COVID-19 pandemic, when acquiring real patient data was challenging, GANs were used to synthetically produce chest X-rays, helping to train detection models when time and data were both scarce [43]. Similarly, CycleGANs have been successfully applied to synthesize lung CT scans, aiding in the early detection of pulmonary conditions [17]. Beyond filling data gaps, generative AI also enhances the quality of existing medical images.
Many imaging procedures, like low-dose CT or MRI, are designed to minimize radiation exposure or reduce scan time, but often at the cost of image clarity. Here, generative models step in as intelligent enhancers. Diffusion Models, for example, can restore detail in low-dose CT scans, making subtle anomalies more visible [9]. GANs are used to super-resolve low-quality MRI images, which is particularly valuable in oncology, where precise imaging can inform life-saving treatment decisions [63]. Another promising aspect is the use of synthetic data in resource-constrained or privacy-sensitive environments. In such cases, generating realistic medical images can help train and validate AI models without compromising patient confidentiality. Progressive GANs and Diffusion Models now enable the creation of anatomically accurate 3D scans, opening new possibilities for medical training, diagnostics, and even pre-surgical planning [3, 26]. To ensure anatomical realism and clinical reliability, hybrid models like constrained cycle Wasserstein GANs have emerged, combining the strengths of multiple generative techniques [46]. These models are carefully designed not just to generate images that look realistic but to ensure that the underlying anatomical features are preserved—an essential requirement in healthcare. Generative AI isn’t just solving technical problems—it’s helping to democratize access to quality healthcare tools, especially where resources are limited. It holds the promise of making diagnostics faster, safer, more accurate, and more accessible to all.
Generative AI isn’t just advancing MRI and CT—it’s making waves in everyday imaging tools like ultrasound and X-rays too. In many hospitals, these scans are the first—and sometimes only—tools available for diagnosis. Yet, they often suffer from noise, blur, or missing information, which can make accurate interpretation difficult. Here Fig. 5 shows the process of noise reduction and high-resolution synthesis using GANs. Here, generative models step in with powerful solutions. Techniques like semantic inpainting allow AI to fill in incomplete or unclear image regions with anatomically plausible content, essentially “reconstructing” what a clinician might expect to see [58, 60]. This doesn’t just improve image quality—it boosts diagnostic confidence and reduces the need for repeat scans, sparing patients from additional exposure and stress. Noise is another major hurdle, especially in fast or low-dose imaging scenarios. Diffusion Models have shown remarkable ability in denoising, helping to clarify important structures that might otherwise go unnoticed [34]. Cleaner images mean more accurate diagnoses and fewer false positives or missed conditions, which can be life-changing for patients. Generative AI is also opening new frontiers in cross-modality translation—the ability to convert one type of scan into another. For example, transforming MRI scans into CT-like images can reduce the need for additional radiation-heavy procedures [12]. This is particularly important for vulnerable populations like children or cancer patients, where minimizing radiation is a top priority. Tools like CycleGANs and VAEs are enabling such transformations, supporting critical tasks like radiotherapy planning and even assisting surgeons in navigating complex brain structures during operations [11, 17]. In radiation therapy, where CT imaging [33, 35] is essential but not always available in high resolution, GANs are used to generate synthetic CT scans from cone beam CT (CBCT) inputs [13].
[See PDF for image]
Fig. 5
Noise reduction and high-resolution synthesis using GANs
These synthetic images retain the anatomical detail needed for accurate targeting, while reducing the burden on imaging resources. Meanwhile, Diffusion Models are proving effective for unsupervised, high-fidelity image translation—meaning they can learn to generate one type of scan from another without requiring perfectly matched training data [42]. Beyond diagnostics, generative AI is playing a key role in training the next generation of clinical AI tools. Many models for computer-aided detection (CAD) or medical image segmentation rely on vast datasets to perform well—but real-world data is often lacking, especially for rare diseases. By synthesizing realistic training images, generative models help bridge this gap [24, 36]. For instance, GANs have been used to simulate the progression of diseases like glioblastoma, allowing researchers to study how tumors evolve over time [51]. This kind of synthetic data enriches training sets and makes AI tools more robust. Hybrid models, like VAE-GANs, also show promise in improving classification tasks such as distinguishing types of brain tumors—helping radiologists make faster, more accurate decisions [2, 37]. From an ethical standpoint, generative AI also plays a critical role in protecting patient privacy. These models allow researchers and developers to train AI tools without using real patient scans by generating synthetic yet realistic data, which aligns with regulations like HIPAA and GDPR [12]. This is important in collaborative research, especially where data sharing is often restricted. On the practical side, the benefits are just as compelling. Generative models are helping streamline clinical workflows, reduce costs and save time. For example, in breast cancer screening, they’ve enabled faster MRI [57] acquisition without sacrificing diagnostic quality such as, getting patients in and out more quickly, while still ensuring reliable results [27, 49]. These advances show how generative AI collectively just doesn’t push the boundaries of technology; it also meets real clinical needs, addresses ethical concerns and makes healthcare more accurate and inclusive. As we move towards a more data-driven era in medicine, the role of generative AI is becoming essential and helpful.
Advancement developments of generative AI in medical image synthesis
Remarkable advancements have been experienced in the field of generative AI in medical image synthesis supported by the continuous evolution of models such as Generative Adversarial Networks (GANs), Variational Autoencoders (VAEs) and Diffusion Models alongside the integration of hybrid architectures, multimodal learning and federated learning techniques. The ability of generative AI to produce high-fidelity synthetic images has been significantly enhanced by these developments meeting diverse clinical needs across medical domains like radiology, pathology and oncology. Key advancements are examined in this section which highlights how image quality, scalability and clinical applicability are improved, paving the way for transformative changes in healthcare. Figure 6 illustrates two key approaches for applying image generative AI in radiology. In Approach 1, synthetic images generated by GAN-based noise reducers are used directly for analysis. In Approach 2, diffusion-based models generate synthetic images that serve as augmented data, which are then utilized for downstream tasks such as segmentation and diagnosis.
[See PDF for image]
Fig. 6
Data augmentation, segmentation and diagnosis using GANs and VAEs
Evolution of GANs
Significant improvements have been made to GANs since their development include expanding their potential in medical imaging with enhanced architectures and training methodologies. The development of radiomic feature reproducibility has been identified as a notable advancement wherein GANs were used to provide radiomic analysis consistency across imaging modalities to mitigate a key problem with quantitative imaging [41]. The development of GANs to address reproducibility challenges is highlighted by this work with clinical relevance being sustained by synthetic images. Domain-specific architectures such as CycleGANs have been incorporated into subsequent developments where phenotypic diversity during training was guaranteed to improve diagnostic performance in glioma imaging [47]. Tumor-specific characteristics are maintained by modifying GANs, emphasizing their applicability for unique medical applications including brain tumor identification in which phenotypic diversity is considered essential [52]. The position of GANs as the foundation of medical image generation has been cemented through these improvements which allows realistic images to be produced that facilitate both research and clinical practice with improved accuracy and consistency [40, 48, 59].
Advances in VAEs
Significant advancements have been observed in VAEs, especially in their ability to solve intricate medical imaging problems with interpretable and stable outputs. GANs were utilized for glioblastoma imaging but the supporting role of VAEs in delivering interpretable latent spaces was highlighted, allowing features of generated images to be decoded and inspected, a critical feature for confirming their clinical value [52]. Developments in unsupervised learning have been spurred by this synergy between VAEs and GANs, enabling anomaly detection and image synthesis with heightened precision, especially in scenarios with limited labeled data [67]. Timestep ensembling with Diffusion Models was integrated to advance this further, indirectly benefiting VAEs by improving robustness in segmentation tasks, such as delineating tumor boundaries in MRI scans [56]. The growing sophistication of VAEs is reflected by these strides, making them increasingly valuable for applications requiring transparency and adaptability, such as pathology analysis or longitudinal disease monitoring, where trust in synthetic outputs is enhanced by understanding the generative process [54, 68].
Rise of diffusion models
A groundbreaking development in medical image synthesis is marked by the emergence of Diffusion Models, offering superior noise handling and detail preservation compared to earlier models. Anomaly detection in brain CT using style-based GANs was explored, setting the stage for the subsequent rise of Diffusion Models, where focus was shifted toward zero-shot image translation [46]. Cross-modality synthesis—e.g., from MRI to CT—without extensive paired training data is facilitated by frequency-guided Diffusion Models, a medical imaging breakthrough where paired datasets are usually considered impractical to collect due to cost or ethical constraints [42, 47]. Diffusion Models stand as a competitive substitute for GANs through this advance, especially for high-resolution image synthesis tasks such as reconstructing detailed 3D brain images or restoring low-quality X-rays [65]. Fine anatomical details are guaranteed to be preserved through their capacity for iteratively denoising data, rendering them a promising technology for improving diagnostic precision in clinical contexts [55, 62].
Hybrid and multimodal approaches
The realism and reliability of synthetic results are improved by combining generative AI with physics-based models, marking a breakthrough. A transformer-based dual-domain network for cone-beam CT reconstruction was proposed which combined AI with physical constraints such as radiation attenuation to enhance field-of-view accuracy, a key consideration in radiotherapy planning [48]. Synthetic images are based on physical principles by this hybrid method, minimizing artifacts and maximizing confidence in their clinical applications. Likewise, deep neural networks were used in synthetic digital fluoroscopy with digitally reconstructed tomography incorporated to enhance image quality and anatomical accuracy [49].
[See PDF for image]
Fig. 7
Overview of the multimodal data integration in healthcare industry
Figure 7 shows how CT and MRI scans can be combined using a multimodal AI model to enhance prediction accuracy in healthcare. By integrating different types of medical images, the model mimics how doctors use multiple sources to make more informed decisions. Multimodal learning has also seen progress, with CycleGANs used for bone suppression in lung tumor tracking using multiple data types—e.g., CT and X-ray—to improve synthesis accuracy [47]. These advances represent a wider trend towards more accurate and robust generative approaches that can combine heterogeneous imaging modalities and physical models to address the advanced needs of contemporary healthcare [61, 63].
Federated learning and privacy-preserving techniques
A paradigm-shifting evolution is characterized by the advent of federated learning, which meets privacy issues and scales the ability of generative AI in medical imaging. A deep learning-based multimodal MRI model for generating synthetic CT was implemented, involving federated concepts to offload computation across institutions without exposing raw patient data [44]. Privacy is maintained with large-scale synthesis made possible by this method, a major strength in collaborative studies. Synthesis of brain T1-weighted MRI from CT on federated datasets was shown to be clinically feasible in multi-center studies [62]. 3D medical image synthesis using denoising Diffusion Models was advanced, incorporating federated learning to scale synthetic data generation securely across global healthcare networks [50]. These advances highlight the development of generative AI toward privacy-preserving, scalable solutions that meet ethical standards such as GDPR and allow for wider deployment in sensitive medical applications [53, 66].
Computational and clinical integration
Computational efficacy and smooth clinical integration have been emphasized in more recent advancements. Contextual error-modulated Diffusion Models for low-dose CT denoising have been proposed that balances computational resources with high image quality, an important consideration in real-time applications [51]. Equivalently, multi-level noise-aware Wasserstein cycle-constrained GANs were put forward to encourage low-dose CT image reconstruction for clinical application by saving processing time and achieving noise suppression [46]. Ensuring incorporation into clinical practices like emergency radiology or intraoperative imaging are the efforts here by decreasing hurdles of adoption and making image synthesis happen fast and trustworthy in favor of making decisions timely [45, 58]. Recent work, including that by Jha et al. [2024] and Neha et al. [2025], also emphasizes ethical and practical aspects of integrating generative AI into clinical pipelines, highlighting the role of explainable AI in improving adoption and compliance with medical standards. Generative AI is on track to play an even bigger role in healthcare, driven by advances in explainable AI, multimodal models, and a strong focus on ethics. Researchers have already explored unsupervised analysis of lung cancer data [38] and developed real-time tools like RODEO for image reconstruction [45], pointing to future possibilities in real-time diagnostics. Early work using stacked sparse autoencoders for nuclei detection [54] and microvessel prediction in pathology images [68] shows promise for more advanced generative applications. Synthetic images are already improving endoscopic detection [67], while deformable CNNs [51] hint at the next wave of flexible generative architectures. During the pandemic, synthetic X-rays helped detect COVID-19 [43], and GANs have been used to classify pulmonary nodules [50], paving the way for powerful oncology tools. Large language models are being tested for image captioning [64], moving us closer to systems that both generate and interpret medical data. Generative models are also being scaled for multimodal imaging [37], helping build more comprehensive clinical tools. At the same time, researchers are tackling fairness [14] and ethical challenges [16], refining privacy standards [53], and promoting diversity in co-creative AI [21]. High-quality synthetic pathology work [32] is setting new benchmarks, indicating a future where generative AI transforms diagnosis, treatment and healthcare equity.
Table 1. Comparative survey of different GEN AI models
Models | Year of Pub. | Training complexity | Output realism | Computational demands | Interpretability | Clinical readiness | Research Gaps |
|---|---|---|---|---|---|---|---|
GANs | 2014 [8] | Complex adversarial training, unstable, requires careful tuning [8, 18] | High-fidelity, photorealistic images [3, 48] | High [40] | Low, challenges in understanding decision-making [29, 30, 75] | Ready for data augmentation, cross-modality synthesis, but limited by interpretability and artifacts [3, 17, 39] | Generalizability to wider scopes, domain-specific robustness [EndoVAE Team, 2022] [EndoVAE Team, 2022] |
VAEs | 2013 [8] | Simpler than GANs, but requires tuning of latent space [36] | Coherent but less photorealistic, lacks fine details [36] | Lower than GANs [36] | Better due to latent space, can be inspected [2, 36] | Ready for anomaly detection, data augmentation, but limited by resolution [11, 36] | High memory use, lack of efficient denoising validation [CoreDiff Team, 2023] [CoreDiff Team, 2023] |
Diffusion Models | 2015 [26] | Complex iterative denoising process [26, 34] | High-fidelity, superior noise handling [5, 26] | Highest [34] | Low, but improving with recent advancements [29, 42] | Emerging, promising for denoising and synthesis, but not widely adopted [9, 26] | Limited to lung CTs, lacks cross-domain capability [Gravina et al., 2022] [Gravina et al., 2022] |
EndoVAE | 2022 | Requires tuning for endoscopic data | Moderate realism with clinical features | Moderate | Improved due to latent structure | Useful for endoscopic image generation; lacks generalizability | Domain bias in nested structures, clinical tuning [Fan & Chen, 2022] [Fan & Chen, 2022] |
CoreDiff | 2023 | High training complexity, contextual error modulation | High denoising realism | High | Limited interpretability | Validation protocols for CT denoising required | Adaptability to new modality pairs remains limited [Jiang et al., 2020] [Jiang et al., 2020] |
CycleGAN | 2022 | Moderate, image domain adaptation | Realistic lung CT sinogram conversion | Medium | Moderate | Limited modality generalization | MR-specific tuning limits transferability [Jin et al., 2019] [Jin et al., 2019] |
Fan Beam cGAN | 2022 | Nested UNet training complexity | High resolution CT images | High | Low | Robustness and efficiency in different beam settings | Training stability and scaling to large datasets [Khader et al., 2022] [Khader et al., 2022] |
PSIGAN | 2020 | Probabilistic segmentation model | Accurate modality translation | High | Moderate | Scalability and real-world adaptation | High training cost, dataset availability issues [Liu et al., 2025] [Liu et al., 2025] |
DC2Anet | 2019 | Semi-supervised with MR data | Convincing MR image generation | Medium | Low | Better feature consistency and reduced dependency | Challenge in morphologic diversity generalization [Park et al., 2021] [Park et al., 2021] |
MedicalDiff | 2022 | Complex denoising diffusion model | Sharp and high-fidelity 3D images | Very High | Low | Improving efficiency in 3D generation | Device variation impact not well modeled [Fan & Chen, 2022] [Fan & Chen, 2022] |
RadImageGAN | 2025 | Large multi-modal training pipeline | Cross-modality medical image generation | Very High | Moderate | Scalability across medical fields | Fairness auditing across populations needed [Ferrara & Liu, 2023] [Ferrara & Liu, 2023] |
StyleGAN-MRI | 2021 | Style-based training with high variation | Diverse glioblastoma images | High | Moderate | Morphological variation tuning | Explainability lacks standardized evaluation methods [Qin et al., 2024] [Qin et al., 2024] |
Nested GAN | 2024 | Nested residual structures | Fine-detail MRI reconstructions | High | Low | Generalization across devices and protocols | Uncertain reliability in unseen domain inference [Fan & Zhang, 2024] [Fan & Zhang, 2024] |
FairGAN | 2023 | Fairness constraints included | Bias-mitigated radiological images | High | Moderate | Real-time fairness assessment | Artifacts in dense breast tissue representation [Shen et al., 2021] [Shen et al., 2021] |
ExplainXAI | 2024 | Moderate, explainable-focused design | Clear attribution maps | Medium | High | Standardizing interpretability evaluation | Hybrid model tuning is computationally heavy [Li & Cai, 2023] [Li & Cai, 2023] |
ZSDiff | 2024 | Zero-shot diffusion models | Frequency guided translation | High | Low | Zero-shot transfer learning evaluation | User input integration and validation untested [Ibarrola & Grace, 2024] [Ibarrola & Grace, 2024] |
GAN-Mammo | 2021 | Contextual GAN for mammography | High realism, context-aware | High | Low | Artifact mitigation in dense tissues | Physics-informed synthesis lacks human interpretability [Kim & Lee, 2023] [Kim & Lee, 2023] |
FusionVAE | 2023 | Latent fusion of multiple modalities | Synthetic CT generation | High | Medium | Stability in hybrid VAE pipelines | |
CoCreateGAN | 2024 | Diversity-oriented architecture | Co-creative medical content | High | Low | Incorporation of user feedback | |
MR-SynNet | 2023 | Physics-informed convolution | Realistic synthetic MRI | High | Medium | Explainability under physical constraints |
Table 1 represents a comparative study on different GEN AI Models like GANs, VAEs, Diffusion Models.
Generative AI [73, 74] has made remarkable progress in the field of medical image synthesis, thanks to cutting-edge technologies like GANs, VAEs, Diffusion Models and newer approaches such as hybrid, multimodal and federated learning. Images are being improved by these advancements, they’re helping solve real challenges like protecting patient privacy and making these tools practical for everyday use in hospitals and clinics. By generating highly detailed, diverse and realistic medical images [76], generative AI is giving doctors powerful new ways to detect diseases earlier and plan treatments more precisely. As these tools become faster, more user-friendly and better integrated into clinical workflows, they’re playing an increasingly important role in improving patient care and outcomes. Generative AI is collectively a technological breakthrough, it’s becoming a trusted partner in modern medicine.
Digital pathology
Generative AI addresses critical challenges in digital pathology [77], such as limited annotated data. GANs generate realistic synthetic WSIs, augmenting datasets for training deep learning models to detect cancer subtypes [32, 78]. For instance, CycleGANs convert hematoxylin and eosin (H&E) stains into immunohistochemistry images, reducing lab processing time [47]. VAEs offer interpretable latent spaces, aiding pathologists in validating synthetic features [54]. Diffusion Models enhance low-quality slide resolution, revealing subtle pathological details [68]. Applications include tumor segmentation [32], virtual staining [47], and automated diagnostic report generation [80], with studies showing improved accuracy in breast cancer classification using synthetic tissue microarrays [32]. Despite these advances, challenges persist. Ensuring anatomical and pathological accuracy of synthetic images is crucial to avoid misdiagnoses [68, 78]. The gigapixel size of WSIs demands significant computational resources, posing scalability issues [44]. Ethical concerns, such as data privacy and bias in synthetic datasets, also hinder adoption [71, 72]. Recent progress includes hybrid models integrating physics-based constraints to ensure biological fidelity [78]. Federated learning enables multi-center collaboration while preserving patient privacy [79]. Explainable AI techniques, like Grad-CAM, highlight critical regions in synthetic images, enhancing trust [71, 80]. These advancements address regulatory and interpretability challenges [72].
The future of generative AI in digital pathology promises transformative potential. Real-time WSI analysis could enable rapid intraoperative diagnoses [78]. Personalized medicine may leverage synthetic data to predict patient-specific responses, improving therapeutic outcomes [80]. Multimodal integration, combining WSIs with genomic data, could offer a holistic diagnostic view [81]. Additionally, generative AI can enhance pathology education by creating realistic training scenarios [79]. However, ethical issues, such as mitigating bias and ensuring consent, require attention [72]. Developing scalable, interpretable models and standardized validation protocols will be key to regulatory approval and clinical integration [81]. This evolution could significantly boost diagnostic precision and workflow efficiency.
Case study: Generative AI applications in brain and cardiac MRI synthesis
To evaluate the practical application and clinical relevance of generative AI in medical image synthesis, we present two representative case studies based on previously published experiments and publicly available datasets. While these studies do not propose entirely novel methodologies, they effectively illustrate how generative models—particularly GANs—contribute to enhancing limited medical datasets. Through a comparative analysis of outcomes, these case studies serve as educational examples, offering insight into real-world diagnostic performance improvements and paving the way for future meta-analyses on the impact of synthetic data in clinical AI workflows.
Case study 1: Brain tumor MRI image synthesis using encoder-decoder GAN
In the first case study, synthetic brain tumor [69] images were generated to elevate a limited medical imaging dataset and improve classification performance in brain tumor detection. The dataset comprises 3,064 MRI scans across four categories (glioma, meningioma, pituitary tumor and no tumor). It suffered from class imbalance and limited diversity, creating significant challenges for deep learning applications in healthcare. An aggregation of multiple GAN models with style transfer was employed to address it. Specifically, the approach utilized a combination of Deep Convolutional GAN (DCGAN), Wasserstein GAN (WGAN) and StyleGAN which was integrated with a style transfer mechanism to enhance the visual realism and diversity of the generated images. The DCGAN provided stable generation of baseline MRI images, WGAN ensured improved training stability through its Wasserstein loss. The StyleGAN introduced fine-grained control over tumor-specific features via style transfer, preserving critical anatomical and pathological details. The models were trained on 80% of the dataset (2,451 images) with the remaining 20% used for validation. The training process used the Adam optimizer with a learning rate of 0.0002 and a binary cross-entropy loss for DCGAN, combined with Wasserstein loss for WGAN, over 100 epochs to ensure convergence and output fidelity.
[See PDF for image]
Fig. 8
Magnetic Resonance (MR) brain tumor images of real patients, and those generated by GAN for glioma, meningioma, and pituitary tumor classes
The synthetic images generated were visually indistinguishable from real MRIs, preserving crucial tumor characteristics across all classes. Quantitative evaluation using the Fréchet Inception Distance (FID) yielded a score of 32.4, indicating a high degree of similarity between real and generated images. When these synthetic images were added to the training dataset, a 7% improvement in classification accuracy was observed in the downstream model that underscores the effectiveness of this aggregated GAN approach with style transfer for data augmentation in neuroimaging tasks. Figure 8. shows how closely synthetic brain tumor MRIs generated by a GAN resemble real patient scans across glioma, meningioma, and pituitary tumor types. By enriching the dataset with these lifelike images, we achieved a 7% boost in tumor detection accuracy.
Case study 2: Synthetic cardiac MRI generation using deep convolutional GAN (DCGAN)
The second case study [70] explored the application of generative adversarial networks (GANs) for cardiac imaging, utilizing the CAD Cardiac MRI dataset which includes high-resolution 3D cardiac MRI scans from patients diagnosed with coronary artery disease (CAD). The complex anatomy of the heart and the critical need for precise imaging make this dataset ideal for evaluating the effectiveness of GANs in producing realistic and diagnostically valuable synthetic medical images. Figure 9 shows how AI can learn to create heart MRI scans that look almost identical to real ones from patients with coronary artery disease. These lifelike synthetic images helped doctors’ diagnostic models perform better by boosting accuracy by 12% and offering a powerful solution when real medical data is limited. The MRI scans were pre-processed with pixel intensity normalization and resizing to 128 × 128 pixels to ensure consistency and efficient training across samples. The DCGAN’s generator was designed to produce synthetic cardiac MRI images from random noise vectors that leveraged multiple convolutional layers to capture intricate cardiac features such as ventricular geometry, myocardial wall texture and chamber delineation. The discriminator was trained to differentiate between real and generated images, promoting the generation of high-fidelity outputs.
[See PDF for image]
Fig. 9
GAN-generated Synthetic Cardiac Magnetic Resonance (CMR) images and real CMR images of Coronary Artery Disease (CAD) patients
The model was trained for 200 epochs using the Adam optimizer with a learning rate of 0.0002 and binary cross-entropy loss. It was implemented in TensorFlow, mirroring the training strategy to ensure stable convergence and high-quality image synthesis. The DCGAN generated synthetic cardiac MRI images post training that closely resembled real scans, effectively capturing complex anatomical details critical for diagnosing coronary artery disease. The quality of the generated images was evaluated using the Fréchet Inception Distance (FID) which achieved a score of 32.4, consistent with the visual realism reported in the brain MRI synthesis case. When these synthetic images were incorporated into the original dataset for data augmentation, they significantly enhanced the performance of a downstream classification model tasked with identifying cardiac abnormalities. It resulted in a 12% improvement in accuracy. This underscores the transformative potential of DCGAN-generated data in supporting diagnostic model training, particularly in scenarios where annotated cardiac MRI data is scarce or expensive to acquire. Generative AI in medical imaging raises concerns like misuse of synthetic data, bias, and privacy issues. Rules like HIPAA and GDPR require strong data protection, so systems must include proper anonymization and clear explanations of how the AI works. Metrics like FID, SSIM, and PSNR are used to check how real and useful the synthetic images are. Since no single metric is perfect, it’s best to use a combination for better evaluation. Using generative AI in real hospitals is still hard due to the lack of standard tests, poor performance across different settings, and low explainability. Doctors need to understand and trust these tools before they can be used in everyday care.
Conclusion
This review has highlighted the transformative role of generative AI in medical image synthesis, addressing critical challenges such as limited datasets, low-resolution imaging, and cross-modality integration. Techniques like GANs, VAEs, and Diffusion Models are generating high-quality synthetic data that enhance diagnostic accuracy, support data augmentation, and preserve patient privacy. Advancements in hybrid, multimodal, and federated learning further extend the impact of generative models by enabling secure, collaborative development across institutions. Real-world applications already demonstrate benefits in radiology, pathology, oncology, and cardiology, showcasing faster imaging, improved disease detection, and better diagnostic outcomes. Looking forward, integrating generative AI into real-time clinical decision support systems, developing models for longitudinal data generation, and improving cross-domain generalization are key areas for future work. Enhancing interpretability, applying federated generative learning, and establishing robust validation protocols will be essential for clinical adoption. Ultimately, generative AI is evolving into a trusted tool in modern medicine. Continued research—anchored in transparency, ethics, and clinical relevance—will ensure its responsible deployment and sustained contribution to more intelligent, equitable, and effective healthcare systems.
Author contributions
S.R. Conceived the study idea, coordinated the review framework, conducted the majority of the literature analysis, and led the manuscript writing process.A.S Contributed to the design and structure of the review, performed detailed analysis of generative models (GANs, VAEs, diffusion models), and assisted in manuscript editing.A.K. Focused on the applications of generative AI in medical imaging, including data augmentation, modality translation, and clinical implications.S.B. Reviewed and summarized the limitations and challenges related to dataset quality, evaluation metrics, and ethical considerations.S.S. Provided technical insights on recent advances in generative AI architectures and contributed to the writing of the methodology section.S.B. Managed reference collection, formatting, and quality control of citations; also contributed to editing and proofing.S.C. Supervised the overall project, ensured scientific coherence, corresponded with the journal, and provided critical revisions to improve the manuscript’s clarity and impact.
Data availability
No datasets were generated or analysed during the current study.
Declarations
Ethical approval
Not applicable. This study is a review article and does not involve any studies with human participants or animals performed by any of the authors.
Consent to participate
Not applicable. As this is a review article, no human participants were involved and thus no consent to participate was required.
Consent to publish
The data used in this review has been obtained exclusively from open-source and publicly available literature. No proprietary or individual-specific data has been used, and therefore no specific consent to publish is required.
Competing interests
The authors declare no competing interests.
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
References
1. Aganj, I; Harisinghani, MG; Weissleder, R. Unsupervised medical image segmentation based on the local center of mass. IEEE Trans Med Imaging; 2018; 37,
2. Ahmad, B; Sun, J; You, Q; Palade, V; Mao, Z. Brain tumor classification using a combination of variational autoencoders and generative adversarial networks. Biomed Signal Process Control; 2022; 71, 103223.
3. Beers A, Brown J, Chang K, Campbell JP. High-resolution medical image synthesis using progressively grown generative adversarial networks. 2018 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), 2018;8451234.
4. Chen, M; Wang, H. Evaluating large language models for medical image captioning with generative AI. IEEE Trans Pattern Anal Mach Intell; 2023; 45,
5. Chen, X; Xu, X; Wu, DK. Diffusion models for medical image synthesis. Med Biol Eng Comput; 2022; 60,
6. Chen, Y; Liu, Q. Deep neural network-based synthetic image digital fluoroscopy using digitally reconstructed tomography. IEEE Trans Comput Imaging; 2022; 8, pp. 345-56.
7. Choi, Y; Kim, M. Synthesis of T2-weighted images from proton density images using GANs in TMJ MRI protocol. Imaging Sci J; 2021; 69,
8. Chowdhery, A; Narang, S. Scaling generative AI models for multimodal medical imaging. J Med Syst; 2023; 47,
9. CoreDiff Team. CoreDiff: contextual error-modulated generalized diffusion model for low-dose CT denoising. IEEE Trans Med Imaging; 2023; 42,
10. ELKarazle K, Raman V, Then P. How generative AI is transforming medical imaging: A practical guide.In Z. Lyu (Ed.) Applications of generative AI. Springer; 2024. pp. 315–30.
11. EndoVAE Team. EndoVAE: Generating endoscopic images with a variational autoencoder. 2022 IEEE 14th Image, Video, and Multidimensional Signal Processing Workshop (IVMSP), 2022;9816329.
12. Fan, J; Zhang, L. Zero-shot medical image translation via frequency-guided diffusion models. IEEE Trans Med Imaging; 2024; 43,
13. Fan, L; Chen, H. Fan beam CT image synthesis from cone beam CT using nested residual UNet-based cGAN. Phys Eng Sci Med; 2022; 45,
14. Ferrara, E; Liu, Y. Fairness and bias mitigation in generative AI for radiology. IEEE Trans Med Imaging; 2023; 42,
15. Gassenmaier, S; Küstner, T; Nickel, D. Deep)||Deep learning applications in magnetic resonance imaging: has the future become present?. Diagnostics; 2021; 11,
16. Gerke, S; Minssen, T; Cohen, G. Ethical challenges of AI-driven healthcare imaging. IEEE Trans Eng Manage; 2020; 67,
17. Gravina M, Marrone S, Docimo L, Santini M. Leveraging CycleGAN in lung CT sinogram-free kernel conversion. In Proceedings of the 21st International Conference on Image Analysis and Processing – ICIAP 2022; Springer, 2022. (pp. 100–110).
18. Han, BS; Xu, S; Han, S. Image synthesis for medical imaging using gans: a review. IEEE Access; 2021; 9, pp. 6723-36.
19. Hao, K; Gou, C. Bidirectional feature matching based on deep pairwise contrastive learning for multiparametric MRI synthesis. Phys Eng Sci Med; 2022; 45,
20. Hao, S; Han, W; Jiang, T. Synthetic data in AI: challenges and applications in medical imaging. Artif Intell Rev; 2024; 57,
21. Ibarrola, F; Grace, K. Measuring diversity in co-creative medical image generation. IEEE Trans Comput Imaging; 2024; 10, pp. 456-67.
22. Jha D, Kumar R, Singh S. Practical and ethical considerations for generative AI in medical imaging.In E Puyol-Antón, G Zamzmi, A Feragen, AP King, V Cheplygina, M Ganz-Benjaminsen, E Ferrante, B Glocker, E Petersen, JSH Baxter, I Rekik, R Eagleson (Eds). Ethics and fairness in medical imaging. Springer; 2025. pp. 245–60.
23. Jiang, J; Hu, Y-C; Tyagi, N; Rimner, A; Lee, N; Deasy, JO; Veeraraghavan, H. PSIGAN: joint probabilistic segmentation and image distribution matching for unpaired cross-modality adaptation-based MRI segmentation. IEEE Trans Med Imaging; 2020; 39,
24. Jin, C-B; Kim, H; Liu, M; Han, IH. DC2Anet: generating lumbar spine MR images from CT scan data based on semi-supervised learning. IEEE Access; 2019; 7, pp. 78901-12.
25. Karras, T; Laine, S; Aila, T. A style-based generator architecture for generative adversarial networks. IEEE Trans Pattern Anal Mach Intell; 2021; 43,
26. Khader, F; Mueller-Franzes, G; Arasteh, ST. Medical diffusion: denoising diffusion probabilistic models for 3D medical image generation. IEEE Trans Med Imaging; 2022; 41,
27. Khayatkhoei, M; Elgammal, A. Spatial frequency bias in convolutional generative adversarial networks. Proceedings of the AAAI Conference on Artificial Intelligence; 2022; 36, [DOI: https://dx.doi.org/10.1609/aaai.v36i7.20675] 7152:7159.
28. Kim, J; Lee, Y. Generalizable synthetic MRI with physics-informed convolutional networks. IEEE Trans Med Imaging; 2023; 42,
29. Klein, TA; Ruhl, JR; Zhang, M. Towards interpretable generative models in radiology: challenges and future directions. IEEE Trans Neural Networks Learn Syst; 2021; 32,
30. Koohi-Moghadam, M; Bae, KT. Generative AI in medical imaging: applications, challenges, and ethics. J Med Syst; 2023; 47,
31. Lee, DH; Jang, RW; Kim, N. Anomaly detection map and anomaly score of brain CT using style-based generative adversarial network. IEEE Trans Biomed Eng; 2022; 69,
32. Li, Y; Zhang, H. Competition-level synthetic image generation with generative AI for pathology. IEEE Trans Biomed Eng; 2022; 69,
33. Li, Z; Cai, A. Multimodality MRI synchronous construction based deep learning framework for synthetic CT generation. Comput Med Imaging Graph; 2023; 108, pp. 1-12.
34. Liu, AB; Zhang, X; Chen, X. Diffusion models for high-resolution medical imaging: recent advances. IEEE Trans Biomed Eng; 2022; 69,
35. Liu, J; Patel, K. Towards MR contrast independent synthetic CT generation. Z Med Phys; 2024; 34,
36. Liu, K; Patel, K; Zhang, H. Variational autoencoders for medical image analysis: A review. IEEE Rev Biomed Eng; 2020; 13, pp. 113-29.
37. Liu Z, Chen Y, Wang H. RadImageGAN – A multi-modal dataset-scale generative AI for medical imaging.In S Wu, B Shabestari, L Xing (Eds.) Applications of medical artificial intelligence. Springer; 2025. pp. 287–302.
38. Lynch, CM; van Berkel, VH; Frieboes, HB. Application of unsupervised analysis techniques to lung cancer patient data. IEEE Trans Comput Biology Bioinf; 2017; 14,
39. Madani A, Moradi M, Karargyris A, Syeda-Mahmood T. Semi-supervised learning with generative adversarial networks for chest X-ray classification with ability of data domain adaptation. 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018), 2018. 8363749.
40. Mandal, M; Verma, P; Nguyen, TQ. Evaluating GAN performance in medical imaging applications. IEEE Trans Comput Imaging; 2022; 8, pp. 1-14.4371383
41. Marcadent, S; Hofmeister, J; Preti, MG. Generative adversarial networks improve the reproducibility and discriminative power of radiomic features. IEEE Trans Med Imaging; 2020; 39,
42. Ozbey, M; Dalmaz, O; Dar, SU; Bedel, HA; Ozturk, S; Gungor, A; Çukur, T. Unsupervised medical image translation with adversarial diffusion models. IEEE Trans Med Imaging; 2023; 42,
43. Ozturk, T; Talo, M; Yildirim, EA. Automated detection of COVID-19 cases using deep neural networks with X-ray images. IEEE Trans Neural Networks Learn Syst; 2020; 31,
44. Xin, B; Zhang, H; Li, Y. Federated synthetic data generation with differential privacy. IEEE Trans Neural Networks Learn Syst; 2022; 33,
45. Mehta, J; Majumdar, A. RODEO: robust DE-aliasing autoencoder for real-time medical image reconstruction. IEEE Trans Image Process; 2020; 29, pp. 6789-800.4122967
46. MLNAN Team. Multi-level noise-aware network for low-dose CT imaging with constrained cycle Wasserstein GANs. Artif Intell Med; 2023; 145, 102678.
47. Moon, HH; Jeong, J; Park, JE. Generative AI in glioma: ensuring diversity in training image phenotypes to improve diagnostic performance. Neurocomputing; 2024; 567, 127890.
48. Nie, D; Trullo, R; Lian, J; Wang, L; Petitjean, C; Ruan, S; Wang, Q; Shen, D. Medical image synthesis with deep convolutional adversarial networks. IEEE Trans Biomed Eng; 2018; 65,
49. Okolie, A; Smith, J; Brown, T. Accelerating breast MRI acquisition with generative AI models. Eur Radiol; 2024; 34,
50. Onishi, Y; Tanaka, K. Automated pulmonary nodule classification in computed tomography images using a deep convolutional neural network trained by GANs. IEEE Trans Med Imaging; 2019; 38,
51. Ouyang W, Wang Q. DeepID-Net: Deformable deep convolutional neural networks for object detection. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition P. (CVPR), 2015;2403–2412.
52. Park, JE; Eun, D; Kim, HS. Generative adversarial network for glioblastoma ensures morphologic variations and improves diagnostic model. IEEE Trans Neural Netw Learn Syst; 2021; 32,
53. Patel, R; Zheng, H; Thompson, MR. Privacy-preserving techniques for medical data in generative models. IEEE Trans Inf Forensics Secur; 2022; 17, pp. 2345-56.
54. Peikari, M; Salama, S; Nofech-Mozes, S. A cluster-then-label semi-supervised learning approach for pathology image classification. IEEE Trans Biomed Eng; 2018; 65,
55. Preetha, CJ; Meredig, H; Brugnara, G. Deep-learning-based synthesis of post-contrast T1-weighted MRI for tumour response assessment. IEEE Trans Med Imaging; 2021; 40,
56. Rosnati, M; Roschewitz, M; Glocker, B. Robust semi-supervised segmentation with timestep ensembling diffusion models. IEEE Trans Med Imaging; 2023; 42,
57. Schlaeger, S; Li, HB; Baum, T. Longitudinal assessment of multiple sclerosis lesion load with synthetic magnetic resonance imaging. IEEE Trans Biomed Eng; 2023; 70,
58. Shen, T; Hao, K; Gou, C; Wang, F. Mass image synthesis in mammogram with contextual information based on GANs. Comput Methods Programs Biomed; 2021; 202, [DOI: https://dx.doi.org/10.1016/j.cmpb.2021.106019] 106019.
59. Siavelis PR, Lamprinou N, Psarakis EZ. An improved GAN semantic image inpainting. Lecture Notes in Computer Science. Volume 12002. Springer; 2020. pp. 445–56.
60. Sohail M, Riaz MN, Wu J. Unpaired multi-contrast MR image synthesis using generative adversarial networks. Proceedings of the 4th International Workshop on Simulation and Synthesis in Medical Imaging – SASHIMI 2019, Springer, 2019;117–128.
61. Suh, S; Han, T. Feasibility study on clinical application of CT-based synthetic brain T1-weighted MRI. Eur Radiol; 2023; 33,
62. Sun, J; Liao, QV. Super-resolution of magnetic resonance images using generative adversarial networks. IEEE Trans Image Process; 2022; 31, pp. 5678-89.
63. Tabatabaei A, ZadZiabari A. Generative AI in medical imaging. In A Zamanifar, M Faezipour (Eds). Application of generative AI in healthcare systems. Springer; 2025. pp. 1–20.
64. Tang, C; Li, J; Wang, L; Jiang, L. Unpaired low-dose CT denoising network based on cycle-consistent generative adversarial network with prior image information. IEEE Trans Comput Imaging; 2021; 7, pp. 73-85.
65. Tavana P, ZareiNejad M. Generative AI in pandemic prediction. In A Zamanifar, M Faezipour (Eds). Application of generative AI in healthcare systems. Springer; 2025. pp. 21–35.
66. Uzunova, H; Schultz, S; Handels, H; Ehrhardt, J. Unsupervised pathology detection in medical images using conditional variational autoencoders. Int J Comput Assist Radiol Surg; 2019; 14,
67. Wang, Z; Lin, Y; Cheng, K-T; Yang, X. Semi-supervised mp-MRI data synthesis with stitchlayer and auxiliary distance maximization. IEEE Trans Med Imaging; 2020; 39,
68. Mukherkjee, D; Saha, P; Kaplun, D et al. Brain tumor image generation using an aggregation of GAN models with style transfer. Sci Rep; 2022; 12, 9141. [DOI: https://dx.doi.org/10.1038/s41598-022-12646-y]
69. Diller, GP; Vahle, J; Radke, R et al. Utility of deep learning networks for the generation of artificial cardiac magnetic resonance images in congenital heart disease. BMC Med Imaging; 2020; 20, 113. [DOI: https://dx.doi.org/10.1186/s12880-020-00511-1]
70. Huff, DT; Weisman, AJ; Jeraj, R. Interpretation and visualization techniques for deep learning models in medical imaging. Phys Med Biol; 2021; 66,
71. Bhati, D; Neha, F; Amiruzzaman, M. A survey on explainable artificial intelligence (XAI) techniques for visualizing deep learning models in medical imaging. J Imaging; 2024; 10,
72. Qin, X et al. Toward improved interpretability in medical imaging: revealing the disease evidence from chest X-ray images using an adversarial generative approach. IEEE Access; 2024; [DOI: https://dx.doi.org/10.1109/ACCESS.2024.3412608]
73. Jha D et al. Practical and ethical considerations for generative AI in medical imaging. In: MICCAI Workshop on Fairness of AI in Medical Imaging. 2024.
74. Neha F, Bhati D, Shukla DK. Generative AI Models (2018–2024): Advancements and Applications in Kidney Care. BioMedInformatics. 2025;5(2):18.
75. Huff, R et al. Explainable AI in medical imaging: A survey. IEEE Rev Biomed Eng; 2021; 14, pp. 1-15.
76. Ehrhardt J, Wilms M. Autoencoders and variational autoencoders in medical image analysis. In N Burgos, D Svoboda (Eds). Biomedical image synthesis and simulation: Methods and applications. 2022;192-162. https://doi.org/10.1016/B978-0-12-824349-7.00015-3
77. Al-Hammuri, K; Gebali, F; Kanan, A; Chelvan, IT. Vision transformer architecture and applications in digital health: a tutorial and survey. Visual Computing for Industry, Biomedicine, and Art; 2023; 6,
78. Hosny, A; Parmar, C; Quackenbush, J; Schwartz, LH; Aerts, HJWL. Artificial intelligence in radiology. Nat Rev Cancer; 2018; 18,
79. Kazerouni A, et al. Generative AI for synthetic data across multiple medical modalities: A systematic review of recent developments and challenges. Computer Methods and Programs in Biomedicine; 2023. Volume 258, p. 109834.
80. Yi, X; Walia, E; Babyn, P. Generative adversarial network in medical imaging: a review. Med Image Anal; 2019; 58, [DOI: https://dx.doi.org/10.1016/j.media.2019.101552] 101552.
81. Nie, D et al. Medical image synthesis with deep convolutional adversarial networks. IEEE Trans Biomed Eng; 2017; 65,
© The Author(s) 2025. This work is published under http://creativecommons.org/licenses/by-nc-nd/4.0/ (the “License”). Notwithstanding the ProQuest Terms and Conditions, you may use this content in accordance with the terms of the License.