MedDistill: Multilevel Distilled Co-Learning for Unsupervised Medical Image Segmentation

Le résumé fourni par la source

Brain tumor segmentation is crucial for clinical diagnosis and treatment planning. Although deep learning has improved performance, it still heavily depends on pixel-level annotations. Recent efforts in weakly supervised learning, such as synthetic tumor generation and few-shot learning reduce this burden, but still struggle with intricate tumor morphologies, illdefined boundaries, and generalization to real tumors. To address these challenges, we propose MedDistill, an unsupervised brain tumor segmentation framework that enables multi-level collaborative learning between visual foundation models and lightweight segmentation networks. Specifically, a pre-trained object detection model is leveraged to generate candidate tumor bounding boxes, which serve as prompts for the vision foundation model to produce high-quality soft pseudo-masks. A lightweight segmentation network is then deployed as a student model to emulate the segmentation capabilities of the foundation model under label-free conditions via knowledge distillation. To further enhance the distillation process, we introduce intermediate feature-level distillation and design an uncertainty-driven loss weight regularizer (ULWR), promoting a dynamic balance of segmentation loss, probability distillation loss, and feature distillation loss to improve training stability and robustness. Experimental results on public and private brain tumor datasets demonstrate that the proposed framework achieves superior segmentation performance under unsupervised settings, outperforming the existing state-of-the-art methods. The code is available at https://github.com/BUCT-IUSRC/Research_MedDistill.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.