Region Contrast Distillation for Medical Image Segmentation

Le résumé fourni par la source

Recent advancements in deep learning have significantly boosted medical image segmentation, but the increasing number of parameters and computations pose challenges for deploying models in resource-constrained environments. To address this issue, we propose an effective structured knowledge transfer framework that uses a well-performing teacher model to train a lightweight student model. Specifically, we present the region comparison distillation(RCD) method, which first encodes the internal information of different semantic regions into various classes of feature vectors with the help of auxiliary labels and then efficiently extracts and conveys the structured information by comparing the similarity of the same kind of feature vectors of the teacher-student model with the difference of different classes of feature vectors. In addition, we improve the traditional logit layer distillation method by dynamically assigning weights to each pixel point when calculating the KL scatter loss, allowing the student model to focus more on regions that are difficult to segment. Extensive experimental results on three medical segmentation datasets, BTCV, Synapse, and Bladder, show that our method significantly improves the performance of lightweight student models and outperforms current state-of-the-art knowledge distillation methods for image segmentation. It is expected to achieve efficient and real-time medical image segmentation on resource-constrained devices.

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