Skip to content

Generative Artificial Intelligence in Medical Imaging: Foundations, Progress, and Clinical Translation

🕒 Published (v1): 2025-08-07 07:58 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

Comprehensive survey of generative AI in medical imaging spanning GANs, VAEs, diffusion models, Transformers, Mamba, autoregressive models, and foundation models, mapped to the full clinical workflow (acquisition → diagnosis → treatment → prognosis). The authors propose a three-tiered evaluation framework (pixel-level fidelity, feature-level realism, clinical-level utility) and catalog open challenges blocking deployment. Foundation models are positioned as the convergent architecture enabling cross-task, cross-modality generalization.

Problem

Despite rapid progress, generative AI for medical imaging lacks: (1) a unified taxonomy across the clinical continuum, (2) standardized multi-level evaluation protocols bridging technical metrics and clinical relevance, and (3) a clear synthesis of how foundation models integrate with and extend purely generative architectures. Deployment is further blocked by domain shift, hallucination risk, high compute, and regulatory uncertainty.

Method

This is a narrative and structured review, not a primary empirical study. The authors organize existing work along two orthogonal axes—model architecture (GAN, VAE, DPM, Transformer, Mamba, AR, foundation model, each with formal objectives, e.g., InfoNCE contrastive loss for foundation models) and clinical phase (acquisition/reconstruction, diagnosis/synthesis, treatment planning/intraoperative navigation, longitudinal prognosis). They then propose a three-level evaluation pyramid: low-level (MSE, PSNR, SSIM, FSIM), mid-level (FID, KID, LPIPS, MedCLIP-score, FVD), and high-level (radiologist expert review, downstream task performance). Public datasets are catalogued (UK Biobank, TCIA, BraTS21, EchoNet-Dynamic, Medtrinity-25M, etc.) and technical/clinical obstacles are systematically analyzed.

Key Contributions

  • Unified taxonomy mapping seven generative model families to four clinical workflow phases with formal mathematical formulations for each architecture.
  • Proposal of a three-tiered hierarchical evaluation framework integrating pixel-level, feature-level, and clinical-level metrics to bridge technical and clinical assessment.
  • Comprehensive survey of public datasets enabling generative medical imaging (radiology, pathology, ophthalmology, multimodal image–text corpora).
  • Explicit synthesis of how foundation models (large-scale contrastive pretraining) converge with generative architectures to enable zero-shot classification, report retrieval, and text-guided synthesis.
  • Structured discussion of translation barriers: generalization under domain shift, hallucination, interpretability, privacy, regulatory compliance, and compute cost.

Results

(All numbers are from cited primary works referenced in this review, not from the review's own experiments.) - Treatment-aware diffusion model for glioma growth simulation improved future tumor prediction accuracy by >16% over baselines using longitudinal MRI + molecular data. - Temporal-Aware Diffusion Model (TADM) reduced brain volume prediction error by 24% compared to conventional baselines in Alzheimer's longitudinal modeling. - Diffusion model for brain volume prediction in Alzheimer's disease improved performance by >20% over prior methods. - Synthetic CT-based deep model predicted hematoma expansion with AUC 0.91, supporting early clinical decision-making in cerebrovascular disease. - Foundation model (MINIM) integrating low-, mid-, and high-level evaluation showed that multi-level assessment is necessary to capture both statistical robustness and clinical utility. - No head-to-head benchmark table across models is provided; quantitative comparisons are drawn from individual cited studies.

Limitations

  • Review scope is necessarily selective; coverage of each subtopic is partial given the field's breadth.
  • Quantitative comparisons across architectures rely on heterogeneous benchmarks from original papers, preventing direct apples-to-apples assessment.
  • The proposed three-tier evaluation framework is conceptual; no reference implementation or standardized codebase is provided.
  • Clinical translation evidence is largely retrospective; prospective validation data are sparse across all cited works.
  • Foundation model section is relatively brief relative to GAN/diffusion coverage, reflecting the nascent state of generative foundation model integration in medical imaging.
  • Hallucination quantification methods (beyond MedCLIP-score) remain underdeveloped for clinical settings.

Relevance to Foundation Models in Medicine

This review directly addresses the convergence of generative AI and medical foundation models, framing contrastively pretrained multimodal architectures (InfoNCE loss, image–text encoders) as enabling zero-shot synthesis, cross-modality translation, and prognostic generalization across tumor types—capabilities that purely discriminative foundation models lack. The three-tier evaluation framework is particularly actionable for researchers benchmarking foundation models, since standard FID/PSNR metrics miss clinical utility. The survey also highlights that foundation models in medicine remain constrained by high training cost, data dependency, and limited interpretability—challenges distinct from those faced by smaller task-specific generative models, pointing to open research directions in efficient pretraining and hallucination control.