What does the strongest evidence actually show?
The most compelling single study is a 2025 dermatology foundation model called PanDerm, trained on over 2 million real-world skin disease images from 11 clinical institutions across 4 imaging modalities [1]. In three reader studies, PanDerm outperformed clinicians by 10.2% in early-stage melanoma detection through longitudinal analysis, improved clinicians' skin cancer diagnostic accuracy by 11% on dermoscopy images, and enhanced non-dermatologist healthcare providers' differential diagnosis by 16.5% across 128 skin conditions on clinical photographs [1]. This is the largest and most comprehensive evaluation among the studies reviewed, covering 28 diverse benchmarks including skin cancer screening, risk stratification, and lesion segmentation [1].
A 2026 prospective study on prostate cancer detection (ProstNFound+) provides another strong piece of evidence: it adapted a medical foundation model for micro-ultrasound and tested it on data collected five years later from a new clinical site [3]. The model showed no performance degradation compared to retrospective evaluation, and its predictions aligned closely with standard clinical scoring protocols (PRI-MUS and PI-RADS) [3]. This is the only prospective validation among the studies, which is a critical step toward real-world readiness [3].
A 2025 osteoporosis screening study using chest X-rays found that a foundation model (DINOv2) achieved an AUC of 0.93 (where 1.0 is perfect) and demonstrated clear clinical reasoning by focusing on relevant bone structures like the spine and ribs [2]. However, this study also revealed an important nuance: medical foundation models did not consistently outperform natural-domain models, and higher performance did not always correlate with better explainability [2]. This suggests that accuracy alone is not enough for clinical trust.
What are the main barriers keeping these models from clinical deployment?
Across all 10 studies, the most frequently cited barrier is the lack of large-scale, prospective clinical validation in real-world settings [4][6][7][9][10]. A 2025 radiology review explicitly states that while foundation models show high performance across tasks, 'a critical gap remains—the translation from research innovation to sustainable clinical radiology practice' [7]. Similarly, a 2026 oncology review notes that 'the scarcity of well-annotated, multi-institutional cohorts limits development and independent validation' [9].
Regulatory and ethical challenges are another major hurdle. A 2025 comprehensive survey of foundation models in medicine highlights that 'despite the transformative potential of FMs, they also pose unique challenges' including algorithmic bias, data privacy, and the need for robust governance mechanisms [10]. A 2026 review on generative AI in healthcare adds that 'regulatory guidelines and policy making for FM-enabled diagnostics are still maturing' [6]. The 2023 chatbot review warns about 'misinformation, inconsistencies, and lack of human-like reasoning abilities' in current models [5].
Explainability—the ability for clinicians to understand why a model made a particular prediction—is a recurring concern. The osteoporosis screening study explicitly found that 'higher performance did not always correlate with better explainability' and argued that 'explainability should be prioritized alongside accuracy in medical AI' [2]. The pathology foundation model review also notes that 'several challenges persist in the clinical application of FMs, which healthcare professionals, as users, must be aware of' [8].
Where is the evidence mixed or incomplete?
The studies disagree on whether medical-specific foundation models are always better than general-purpose ones. The osteoporosis screening study found that 'medical foundation models did not consistently outperform natural-domain models' [2], while the dermatology study showed that a model trained specifically on medical data outperformed clinicians [1]. This discrepancy likely reflects differences in task complexity and data availability—dermatology has abundant, standardized imaging data, while osteoporosis screening from chest X-rays is a more opportunistic, less standardized task.
Another gap is the lack of head-to-head comparisons between different foundation models on the same clinical task. Most studies evaluate a single model or a small set of variants, making it difficult to know which approach is best for a given application [4][7][8]. The 2025 radiology review calls for 'standardization of evaluation methodologies' [7], and the 2025 comprehensive survey echoes this, noting 'insufficient standardization of evaluation methodologies' [10].
Finally, while the dermatology and prostate cancer studies show promising results, they represent only two clinical domains. The pathology review [8] and oncology review [9] both emphasize that most foundation models have been tested on narrow tasks (e.g., cancer diagnosis from a single imaging modality) rather than the full spectrum of clinical decision-making. As the 2025 comprehensive survey puts it, 'a critical gap remains—the translation from research innovation to sustainable clinical radiology practice' [7].
About These Sources
This answer is built on 10 peer-reviewed studies — published from 2023 to 2026, 9 from 2024 or later, 6 in Q1 journals, collectively cited 173 times — selected as the most relevant from 10 studies that passed quality screening, drawn from 58 papers retrieved from a database of over 500 million.
Sources used in this answer
A multimodal vision foundation model for clinical dermatology
In the largest and most comprehensive evaluation among these studies, a dermatology foundation model (PanDerm) trained on over 2 million images from 11 institutions outperformed clinicians by 10.2% in early melanoma detection and improved non-dermatologists' diagnostic accuracy by 16.5% across 128 skin conditions [1].
Explainable opportunistic osteoporosis screening from chest X-rays: a retrospective comparison of foundation models
In a retrospective study of 21,031 chest X-rays, the DINOv2 foundation model achieved an AUC of 0.93 for osteoporosis screening but medical foundation models did not consistently outperform natural-domain models, and higher accuracy did not always mean better explainability [2].
ProstNFound+: A Prospective Study using Medical Foundation Models for Prostate Cancer Detection
In the only prospective validation among these studies, a prostate cancer detection model (ProstNFound+) showed no performance drop when tested on data from a new clinical site five years later, and its predictions aligned with standard clinical scores [3].
Foundation Models in Radiology: What, How, Why, and Why Not
A 2025 radiology review explains that foundation models are trained on large unlabeled data and show high performance across tasks, but emphasizes the need for safe and responsible training to benefit patients and providers [4].
AI Chatbots in Clinical Laboratory Medicine: Foundations and Trends
A 2023 review of AI chatbots in laboratory medicine finds that while chatbots hold potential for education and result interpretation, they suffer from misinformation, inconsistencies, and lack of human-like reasoning, requiring extensive validation before clinical use [5].
Generative AI for Healthcare: Foundation Models, Clinical Applications, Ethical Considerations, and Regulatory Challenges
A 2026 review of generative AI in healthcare identifies key barriers including algorithmic bias, lack of large-scale clinical validation, and immature regulatory frameworks, concluding that successful implementation depends on robust governance and ethical safeguards [6].
Large-Scale Foundation Models for Radiological Image Analysis: Clinical Applications, Technical Challenges, and Future Directions
A 2026 radiology review highlights a critical gap between research innovation and clinical practice, providing deployment strategies and noting challenges in clinical validation, regulatory approval, and ethical implementation across 15 healthcare systems [7].
Pathology Foundation Models
A 2025 pathology review reports that foundation models have been applied to disease diagnosis, survival prediction, and biomarker scoring, but notes persistent challenges in clinical application that healthcare professionals must be aware of [8].
Foundation models in clinical oncology: Progresses and perspectives.
A 2026 oncology review finds that foundation models have advanced screening, diagnosis, and treatment selection, but key barriers include variability in data standards, scarcity of multi-institutional cohorts, and maturing regulatory guidelines [9].
A Comprehensive Survey of Foundation Models in Medicine
A 2025 comprehensive survey of foundation models in medicine covers their evolution, applications across clinical NLP, medical imaging, and omics, and highlights challenges including algorithmic bias, data privacy, and the need for responsible deployment [10].