Generative AI in clinical (2020–2025): a mini-review of applications, emerging trends, and clinical challenges

Generative artificial intelligence (G-AI) has moved from proof-of-concept demonstrations to practical tools that augment radiology, dermatology, genetics, drug discovery, and electronic-health-record analysis. This mini-review synthesizes fifteen studies published between 2020 and 2025 that collectively illustrate three dominant trends: data augmentation for imbalanced or privacy-restricted datasets, automation of expert-intensive tasks such as radiology reporting, and generation of new biomedical knowledge ranging from molecular scaffolds to fairness insights. Image-centric work still dominates, with GANs, diffusion models, and Vision-Language Models expanding limited datasets and accelerating diagnosis. Yet narrative (EHR) and molecular design domains are rapidly catching up. Despite demonstrated accuracy gains, recurring challenges persist: synthetic samples may overlook rare pathologies, large multimodal systems can hallucinate clinical facts, and demographic biases can be amplified. Robust validation, interpretability techniques, and governance frameworks therefore, remain essential before G-AI can be safely embedded in routine care.

Introduction

Healthcare has long grappled with the twin problems of data scarcity and data privacy. Curating large, balanced, and publicly shareable clinical datasets is expensive, logistically complex, and ethically sensitive. Recent advances in generative artificial intelligence (G-AI)—notably Generative Adversarial Networks (GANs), variational auto-encoders, diffusion models, and large Vision-Language Models (VLMs)—offer a potential remedy by synthesising realistic yet privacy-preserving data. Table 1 collates fifteen representative studies that demonstrate how these models are already reshaping diverse clinical tasks.

Table 1

Table 1. Summary of clinical generative AI applications.

Medical imaging remains the most prolific test-bed for G-AI. Early work by Han et al. introduced “pathology-aware” GANs that augment computer-aided-diagnosis (CAD) datasets and serve as training material for novice radiologists (1). Subsequent studies refined both fidelity and dimensionality of synthetic images. Aydin et al. re-engineered StyleGANv2 to generate three-dimensional Time-of-Flight MR angiography volumes, boosting multiclass artery segmentation without additional patient scans (2). Similar philosophies underpin Pawlicka et al.'s colorectal-polyp synthesis, where GAN-generated images alleviate class imbalance and improve endoscopic segmentation accuracy (3). Ultsch and Lötsch addressed melanoma detection by fine-tuning a latent Stable Diffusion model, proving that diffusion-based methods can rival GANs for dermoscopic realism (4).

The promise of G-AI is not limited to raw pixels. Phipps et al. explored VLMs that translate chest x-ray features into free-text radiology reports, potentially reducing radiologist workload during high-volume shifts (5). However, their evaluation framework also revealed a tendency to hallucinate clinical findings—a stark reminder that factual grounding remains a critical bottleneck. Complementary efforts by Huang et al. in emergency-department workflows corroborate both the efficiency gains and the evaluation challenges of text-generating models (6).

Beyond imaging, G-AI is venturing into molecular and systemic domains. Zeng et al. leveraged ProteinGAN and hierarchical generative models to design novel proteins and small molecules, accelerating the pre-clinical discovery pipeline (7). Bordukova et al. harnessed synthetic patient trajectories to construct digital twins that can de-risk costly clinical trials (8). At the intersection of fairness and analytics, Khosravi et al. generated radiographs that isolate race-linked imaging features, providing a sandbox for bias audits (9).

These successes nonetheless surface persistent limitations. Synthetic data often fails to capture rare anatomical variants or subtle disease phenotypes, risking model over-confidence in out-of-distribution scenarios (4, 10). Bias in training corpora can be magnified, as evidenced by demographic skew in pelvic-radiograph synthesis (9). Large multimodal systems may produce credible but incorrect statements, undermining clinical trust (5, 11). Interpretable frameworks such as StylEx, which links StyleGAN latents to human-readable attributes, are therefore gaining traction (12).

Regulatory and ethical considerations further complicate deployment. Frictionless data-sharing enabled by G-AI must still honor patient consent and institutional review protocols. Meanwhile, explainability demands are intensifying; clinicians and regulators alike now expect transparent reasoning pathways before sanctioning AI-assisted decisions. Collectively, the studies surveyed here illuminate both the transformative potential of G-AI and the rigorous safeguards required for its responsible translation to bedside practice.

Methodology of literature selection

To identify relevant studies, we conducted a targeted search in PubMed, IEEE Xplore, and Scopus databases covering January 2020–May 2025. Keywords included “generative AI”, “synthetic data”, “clinical practice”, and “healthcare”. From over 65 initial hits, we prioritised peer-reviewed articles that explicitly applied generative AI in clinical contexts. Fifteen representative studies were chosen to illustrate diverse domains (imaging, text, molecular design, and fairness). These were not intended as an exhaustive list, but rather as exemplars highlighting the breadth and key limitations of generative AI in healthcare.

Comparative analysis and discussion

Table 1 distills fifteen recent studies that deploy generative AI (G-AI) across the clinical data spectrum, with medical imaging emerging as the prime test-bed. More than two-thirds of the entries apply GANs, diffusion models or Vision-Language Models (VLMs) to synthesize, augment or interpret radiographs, MRI volumes and dermoscopic, endoscopic or fundus photographs. These image-centric efforts tackle three chronic bottlenecks highlighted in Table 1: limited data volume, class imbalance and privacy restrictions. For example, Ultsch & Lötsch fine-tune Stable Diffusion to balance melanoma classes, while Aydin et al. extend StyleGANv2 to 3-D angiography volumes, boosting vascular-segmentation accuracy without collecting new scans.

Beyond imaging, Table 1 shows G-AI penetrating narrative and molecular domains. Alkhalaf et al. couple a retrieval-augmented Llama-2 with zero-shot prompting to summarise malnutrition risk factors from electronic health records, illustrating how foundation models can tame unstructured clinical text. Zeng et al. harness ProteinGAN to generate bespoke proteins, signalling a shift from data augmentation to de-novo biomedical design. Meanwhile, Pinaya and Bordukova exploit diffusion models to create synthetic chest x-rays and digital-twin trajectories respectively, lowering the cost and ethical burden of large-scale trials.

The table also exposes recurring limitations. Synthetic samples often omit rare pathologies, risk distribution shifts (e.g., Pawlicka's colorectal polyps) or encode demographic biases (Khosravi's race-aware radiographs). VLMs hallucinate clinical facts, undermining trust in auto-generated reports. Several authors therefore call for stronger interpretability—Lang's StylEx explicitly pairs StyleGAN with attribute visualisation—and for rigorous external validation before clinical rollout.

Collectively, the evidence in Table 1 suggests three near-term pay-offs: (i) privacy-preserving data augmentation that accelerates model development, (ii) automation of expert-intensive tasks such as radiology reporting or phenotype annotation, and (iii) exploratory insight generation that surfaces novel biomarkers or inequities. Realising these benefits, however, hinges on closing interpretability gaps, curbing bias propagation, and establishing governance frameworks that keep pace with rapidly evolving G-AI toolchains. To mitigate these concerns, safeguards such as bias audits, explainability techniques, and transparent provenance tracking of synthetic data should be incorporated into deployment frameworks. Evaluation of generative models is often benchmarked with metrics such as BLEU/ROUGE for text, Fréchet Inception Distance (FID) or Inception Score for images, and perplexity for language models, which provide quantitative grounding for reliability assessments.

Conclusion

Generative AI is already enriching clinical data pipelines, from radiology suites to drug-discovery labs. The reviewed literature confirms tangible gains in diagnostic accuracy, workflow efficiency, and hypothesis generation, driven chiefly by image-focused GANs, diffusion models, and emerging VLMs. Yet every advantage is tempered by unresolved issues of bias, fidelity, and interpretability. Rare pathologies remain under-represented, demographic disparities can be inadvertently reinforced, and text generators are prone to clinically dangerous hallucinations. Future work must therefore pair technical innovation with stringent validation on external cohorts, transparent reporting of synthetic-data provenance, and user-friendly explanation interfaces. Only through such multidisciplinary vigilance can G-AI move from promising prototypes to trustworthy, equity-focused tools that genuinely advance patient care. Emerging trends such as text-to-3D generation for surgical planning signal new directions for generative AI in clinical practice, while broader applications in education and management remain outside the scope of this review.

Author contributions

NF: Software, Investigation, Writing – original draft, Formal analysis, Resources, Writing – review & editing, Funding acquisition, Data curation, Visualization, Validation, Project administration, Conceptualization, Supervision. RR: Data curation, Methodology, Project administration, Validation, Resources, Writing – original draft. SB: Writing – original draft, Conceptualization, Resources. FS: Data curation, Methodology, Conceptualization, Writing – original draft. RA: Conceptualization, Writing – review & editing, Resources, Writing – original draft. FA: Writing – original draft, Resources, Writing – review & editing, Conceptualization. MH: Data curation, Supervision, Conceptualization, Funding acquisition, Writing – original draft, Writing – review & editing. TL: Data curation, Methodology, Writing – original draft, Writing – review & editing. MS: Data curation, Supervision, Formal analysis, Writing – review & editing. KO: Data curation, Formal analysis, Visualization, Writing – review & editing.

Funding

The author(s) declare that no financial support was received for the research and/or publication of this article.

Acknowledgements

The authors want to thank Multimedia University.

Conflict of interest

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The author(s) declare that no Generative AI was used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

References

1. Han C, Rundo L, Murao K, Nemoto T, Nakayama H. Bridging the gap between AI and healthcare sides: towards developing clinically relevant AI-powered diagnosis systems. In: Maglogiannis I, Iliadis L, Pimenidis E, editors. Artificial Intelligence Applications and Innovations. AIAI 2020. IFIP Advances in Information and Communication Technology, vol 584. Cham: Springer (2020). p. 320–33. doi: 10.1007/978-3-030-49186-4_27

Crossref Full Text | Google Scholar

2. Aydin OU, Hilbert A, Koch A, Lohrke F, Rieger J, Tanioka S, et al. Generative modeling of the circle of willis using 3D-StyleGAN. Neuroimage. (2024) 304:120936. doi: 10.1016/j.neuroimage.2024.120936

PubMed Abstract | Crossref Full Text | Google Scholar

3. Pawlicka A, Pawlicki M, Jaroszewska-Choras D, Kozik R, Choras M. Enhancing clinical trust: the role of AI explainability in transforming healthcare. IEEE International Conference on Data Mining Workshops, ICDMW; IEEE Computer Society (2024). p. 543–9. doi: 10.1109/ICDMW65004.2024.00075

Crossref Full Text | Google Scholar

4. Ultsch A, Lötsch J. Augmenting small biomedical datasets using generative AI methods based on self-organizing neural networks. Brief Bioinform. (2024) 26(1). doi: 10.1093/bib/bbae640

PubMed Abstract | Crossref Full Text | Google Scholar

5. Phipps B, Hadoux X, Sheng B, Campbell JP, Liu TA, Keane PA, et al. AI Image generation technology in ophthalmology: use, misuse and future applications. Prog Retinal Eye Res. (2025) 106:101353. doi: 10.1016/j.preteyeres.2025.101353

Crossref Full Text | Google Scholar

6. Huang J, Neill L, Wittbrodt M, Melnick D, Klug M, Thompson M, et al. Generative artificial intelligence for chest radiograph interpretation in the emergency department. JAMA Netw Open. (2023) 6(10):e2336100. doi: 10.1001/jamanetworkopen.2023.36100

PubMed Abstract | Crossref Full Text | Google Scholar

7. Zeng X, Wang F, Luo Y, Kang SG, Tang J, Lightstone FC, et al. Deep generative molecular design reshapes drug discovery. Cell Rep Med. (2022) 3(12):1–13. doi: 10.1016/j.xcrm.2022.100794

Crossref Full Text | Google Scholar

8. Bordukova M, Makarov N, Rodriguez-Esteban R, Schmich F, Menden MP. Generative artificial intelligence empowers digital twins in drug discovery and clinical trials. Expert Opin Drug Discov. (2024) 19(1):33–42. doi: 10.1080/17460441.2023.2273839

PubMed Abstract | Crossref Full Text | Google Scholar

9. Khosravi B, Rouzrokh P, Erickson BJ, Garner HW, Wenger DE, Taunton MJ, et al. Analyzing racial differences in imaging joint replacement registries using generative artificial intelligence: advancing orthopaedic data equity. Arthroplast Today. (2024) 29:101503. doi: 10.1016/j.artd.2024.101503

PubMed Abstract | Crossref Full Text | Google Scholar

10. Pinaya WHL, Graham MS, Kerfoot E, Tudosiu P-D, Dafflon J, Fernandez V, et al. Generative AI for medical imaging: extending the MONAI framework. arXiv preprint arXiv:2307.15208 (2023).

Google Scholar

11. Alkhalaf M, Yu P, Yin M, Deng C. Applying generative AI with retrieval augmented generation to summarize and extract key clinical information from electronic health records. J Biomed Inform. (2024) 156. doi: 10.1016/j.jbi.2024.104662

PubMed Abstract | Crossref Full Text | Google Scholar

12. Lang O, Yaya-Stupp D, Traynis I, Cole-Lewis H, Bennett CR, Lyles CR, et al. Using generative AI to investigate medical imagery models and datasets. EBioMedicine. (2024) 102:1–14. doi: 10.1016/j.ebiom.2024.105075

Crossref Full Text | Google Scholar

13. Bhatt S, Sharma S. Generative artificial intelligence based biomedical applications for pharmaceutical industry. 2025 International Conference on Computational, Communication and Information Technology (ICCCIT); IEEE (2025). p. 1–6

Google Scholar

14. Patel T, Othman AA, Sümer Ö, Hellman F, Krawitz P, André E, et al. Approximating facial expression effects on diagnostic accuracy via generative AI in medical genetics. Bioinformatics. (2024) 40(Supplement_1):i110–8. doi: 10.1093/bioinformatics/btae239

PubMed Abstract | Crossref Full Text | Google Scholar

15. La Salvia M, Torti E, Leon R, Fabelo H, Ortega S, Martinez-Vega B, et al. Deep convolutional generative adversarial networks to enhance artificial intelligence in healthcare: a skin cancer application. Sensors. (2022) 22(16):6145. doi: 10.3390/s22166145

PubMed Abstract | Crossref Full Text | Google Scholar

16. Liu F, Zhou H, Wang K, Yu Y, Gao Y, Sun Z, et al. MetaGP: a generative foundation model integrating electronic health records and multimodal imaging for addressing unmet clinical needs. Cell Rep Med. (2025) 6(4). doi: 10.1016/j.xcrm.2025.102056

Crossref Full Text | Google Scholar