Efficient Medical Segmentation Anything Model for Robust Lightweight Segmentation under Imperfect Data in Embedded Healthcare.
Authors
Abstract
Healthcare Industry 5.0 is accelerating the deployment of medical imaging intelligence on embedded and edge devices, where medical data are often imperfect due to acquisition noise, low-contrast regions, and boundary ambiguity, while computational budgets remain tight. Although the Segment Anything Model (SAM) provides promptable segmentation, its heavy encoder hinders edge deployment and its zero-shot behavior can be unreliable for specialized medical imaging scenarios. This paper presents an efficient and lightweight medical image segmentation pipeline tailored for embedded healthcare systems. We replace the SAM image encoder with an Efficient Vision Transformer (EfficientViT) backbone and introduce offline knowledge distillation to transfer domain-aware feature representations from the Medical Segment Anything Model (MedSAM) to a compact student network. To emulate imperfect supervision in practical use, bounding-box prompts are generated from ground truth masks with random per-side perturbations of up to 20 pixels, reflecting the coarse, time-constrained box prompts typical of bedside and field workflows. In addition, an uncertainty-based refinement strategy is incorporated to improve boundary consistency in ambiguous regions. We train and evaluate the proposed approach on 173,024 image-mask pairs spanning seven medical imaging modalities, using the Dice Similarity Coefficient (DSC), the 95th percentile Hausdorff Distance (HD$_{95}$), and the Normalized Surface Dice (NSD). Experimental results show that the lightweight model achieves segmentation accuracy within 0.02 DSC of MedSAM while reducing inference latency to below one-eighth of MedSAM. Robustness is assessed across multiple prompt-perturbation magnitudes ($r\in \lbrace 0,10,20\rbrace$ pixels) and multiple additive Gaussian noise levels ($\sigma \in \lbrace 0,0.05,0.1,0.2\rbrace$, with $\sigma {=}0$ the clean reference), where the proposed model exhibits stable performance under imperfect inputs and inexact supervision, supporting practical deployment in real-time Internet of Medical Things (IoMT) applications.