Efficient Medical Segmentation Anything Model for Robust Lightweight Segmentation under Imperfect Data in Embedded...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Embedded Systems

Background:

• Healthcare Industry 5.0 drives medical imaging intelligence on edge devices.
• Medical data on edge devices are often imperfect (noise, low contrast, ambiguity).
• Existing models like SAM are too large for edge deployment and unreliable for specialized medical tasks.

Purpose of the Study:

• To develop an efficient and lightweight medical image segmentation pipeline for embedded healthcare systems.
• To adapt large-scale segmentation models for resource-constrained edge devices.
• To ensure reliable segmentation performance under imperfect data and supervision conditions.

Main Methods:

• Replaced SAM's heavy encoder with an Efficient Vision Transformer (EfficientViT).
• Employed offline knowledge distillation from MedSAM to a compact student network.
• Introduced bounding-box prompts with random perturbations and an uncertainty-based refinement strategy.

Main Results:

• Achieved segmentation accuracy within 0.02 DSC of MedSAM.
• Reduced inference latency to less than one-eighth of MedSAM.
• Demonstrated stable performance across various noise levels and prompt perturbations.

Conclusions:

• The proposed lightweight pipeline is suitable for real-time Internet of Medical Things (IoMT) applications.
• The model offers a practical solution for deploying advanced medical imaging intelligence on edge devices.
• The approach effectively handles imperfect medical data and coarse supervision, enhancing clinical workflow efficiency.