Medical imaging foundation models lack a universally superior adaptation strategy

A structured review of 705 independent studies found no consistently preferred way to adapt foundation models for medical image analysis. Across controlled comparisons, frozen, parameter-efficient, partial and full model updating did not receive a stable ranking, and the review also identified no clear winner among broader adaptation families. Performance varied with the task, data regime, comparator and evaluation setting, suggesting that adaptation is a context-dependent design rather than a hierarchy of techniques.

The review, published in Artificial Intelligence Review, focused specifically on post-pretraining interventions applied to existing foundation models. Its systematic, auditable workflow identified 719 eligible reports representing 705 independent studies. Of those reports, 473 provided controlled comparative evidence. The primary synthesis comprised 322 formally published studies, while 151 preprint-only studies formed a separate sensitivity layer. The authors classified adaptation into five non-mutually exclusive families: parameter-level, representation-level, objective-level, data-centric, and architectural or sequence-level adaptation. Model use and update regimes were analyzed separately.

Clinical-readiness evidence was notably limited. Only 173 of the 705 studies, or 24.5%, had confirmed independent external validation, while 10 studies, or 1.4%, reported calibration or uncertainty evidence. Among the 322 formally published comparative studies, 97, or 30.1%, included a qualifying externally validated arm. These results describe the coverage and composition of the available evidence; they do not show that adaptation is ineffective or that any particular clinical application is ready for use.

The central finding is that adaptation outcomes depend on the conditions in which they are assessed. A method that performs well in one comparison cannot therefore be treated as universally superior, and evidence from a narrowly defined task or evaluation setting does not establish the same performance under different conditions. The review is also bounded by its focus on interventions applied after pretraining and by the separation of formally published research from the preprint sensitivity layer. Overall, the available evidence portrays foundation-model adaptation for medical imaging as a design space whose performance and clinical readiness must be evaluated in context, with independent validation, calibration and uncertainty evidence remaining uncommon across the reviewed literature.

Subject of Research: Post-pretraining adaptation strategies and clinical-readiness evidence for foundation models in medical image analysis

Article Title: Adaptation of foundation models in medical image analysis: mechanisms, evidence, and deployment considerations

Article References: Phuntsho, K., Abdullah, Lee, K., Lee, I., & Ahn, E. (2026). Adaptation of foundation models in medical image analysis: mechanisms, evidence, and deployment considerations. Artificial Intelligence Review. https://doi.org/10.1007/s10462-026-11733-0

Image Credits: AI Generated

DOI: 10.1007/s10462-026-11733-0

Keywords: medical imaging, foundation models, model adaptation, external validation, calibration, uncertainty, clinical readiness, systematic review, artificial intelligence, Adaptation, foundation, models