Self-supervised Instance Segmentation of Diabetic Foot Ulcers via Feature Correspondence Distillation

Abstract

Diabetic foot ulcers (DFUs) are a serious complication of diabetes that can often lead to infection, amputation, and even death if not properly managed. Accurate segmentation of DFUs in medical images is crucial for effective treatment planning. In the DFUC2024 challenge, which emphasizes self-supervised learning techniques for DFU segmentation, we investigate two approaches. The first approach utilizes a DINO (self-distillation with no labels) model combined with a trainable clustering probe to map unsupervised features into discrete segmentation labels. The second approach involves modifying the STEGO model, specifically designed to distill unsupervised features into meaningful segmentation labels, by integrating self-attention features from the ViT backbone to enhance spatial information. To further improve segmentation accuracy, we propose a coarse-to-fine instance prediction framework, where initial coarse predictions are refined through focused reprocessing of detected ulcer regions. After optimizing the hyperparameters for the DFU dataset, the modified STEGO model achieves a Dice coefficient of 0.4362 and Jaccard coefficient of 0.3358. Although the proposed approach yields competitive results, the challenge of self-supervision in DFU segmentation remains significant. The implementation of this work is available at https://github.com/Wenhui-Zhang-5/DFUC2024-challenge.