A Light-weight Universal Medical Segmentation Network for Laptops Based on Knowledge Distillation

SONGXIAO YANG, Yizhou Li, Ye Chen, Zhuofeng Wu, Masatoshi Okutomi

Keywords: Segment Anything, Lightweight Model, Medical Imaging Segmentation, Computational Efficiency

Abstract: In medical imaging, accurate and efficient segmentation is crucial for diagnostics, treatment planning, and monitoring disease progression. Traditional methods, while capable of providing reliable results, often require substantial computational resources, which may not be feasible on devices with limited capabilities such as standard CPUs and limited RAM. To address this challenge, we present an optimized universal segmentation framework that leverages a lightweight image encoder RepViT-M0.6, distilled from Swin-T. Our extensive evaluations of the online validation set demonstrate that our approach outperforms the baseline LiteMedSAM model, showing improvements in the Dice Similarity Coefficient (DSC) by approximately 2\% and in the Normalized Surface Dice (NSD) by around 1\%. Furthermore, the method achieves a more than threefold increase in inference speed, making it viable for real-time applications on devices with limited computational power. This demonstrates that our adaptation significantly enhances processing speed and resource efficiency without sacrificing accuracy.

Submission Number: 2