Segment Anything Models for medical image segmentation: From native capabilities to clinical requirements.
Authors
Affiliations (5)
Affiliations (5)
- School of Artificial Intelligence, Chongqing University of Technology, Chongqing, 401135, China. Electronic address: [email protected].
- School of Artificial Intelligence, Chongqing University of Technology, Chongqing, 401135, China.
- School of Smart Health, Chongqing Polytechnic University of Electronic Technology, Chongqing, China.
- Department of Radiology, Daping Hospital, Army Medical University, Chongqing, 400042, China.
- Department of Electrical and Computer Engineering, University of Nevada Las Vegas, Las Vegas, NV, United States.
Abstract
Medical image segmentation is essential for computer-aided diagnosis and precision medicine, yet conventional supervised models depend heavily on high-quality annotations and often generalize poorly across domains. The Segment Anything Model (SAM) family offers a transferable, interactive, and extensible foundation for segmentation, but direct medical transfer remains limited by domain gaps, insufficient modeling of medical semantics and anatomy, and mismatches in high-dimensional volumetric and sequential imaging scenarios. This review critically examines the evolution and native capability boundaries of SAM, SAM 2, and SAM 3. Across generations, native capabilities expand from spatial prompting to sequential propagation and concept-level target specification, without automatically conferring medical expertise. We organize medical adaptation from three connected perspectives: model adaptation and strategy selection, task knowledge organization and utilization, and application-oriented adaptation. Cross-method analysis shows that no strategy is universally optimal; effectiveness depends on the gap between foundation-model capabilities and target-task requirements, together with data scale, domain discrepancy, imaging dimensionality, computational resources, and clinical constraints. We further identify four priorities for clinical translation: standardized evaluation for cross-study comparability, controlled continual adaptation under clinical domain drift, dynamic resource scheduling for high-dimensional and real-time workflows, and uncertainty stratification and clinical reliability control. Overall, further progress depends on accurate capability assessment, condition-matched adaptation, and explicit control of computational and prediction risks.