Abstract
Medical image segmentation has undergone significant changes in recent years, mainly due to the development of base models. The introduction of the Segment Anything Model (SAM) represents a major shift from task-specific architectures to universal architectures. This review discusses the adaptation of SAM in medical visualisation, focusing on three primary domains. Firstly, multimodal fusion frameworks implement semantic alignment of heterogeneous visual methods. Secondly, volumetric extensions transition from slice-based processing to native 3D spatial reasoning with architectures such as SAM3D, ProtoSAM-3D, and VISTA3D. Thirdly, uncertainty-aware architectures integrate probabilistic calibration for clinical interpretability, as illustrated by the SAM-U and E-Bayes SAM models. A comparative analysis reveals that SAM derivatives with effective parameters achieve Dice coefficients of 81-95%, while concomitantly reducing annotation requirements by 56-73%. Future research directions include incorporating adaptive domain hints, Bayesian self-correction mechanisms, and unified volumetric frameworks to enable autonomous generalisation across diverse medical imaging contexts.