Scholarlink

As part of the Diabetic Foot Ulcer Challenge (DFUC) 2024, we propose a novel self-supervised learning (SSL) approach for the segmentation of Diabetic Foot Ulcers (DFUs), leveraging a multi-scale feature extraction architecture combined with the DINOv2 model. This innovative approach is designed to address the variability and complexity in DFU images, which are often characterized by diverse shapes, sizes, and textures that pose significant challenges to traditional segmentation methods. Our method integrates an Attention U-NET as the segmentation head, enabling precise capture of critical ulcer features within DFU images, such as the boundaries and internal structures of ulcers. Using test datasets evaluated by the DFUC 2024 organizers, the model achieved a Dice coefficient of 0.5709 and a Jaccard index of 0.4793, demonstrating its instance segmentation capabilities on unseen data. The relevant code is released at: https://github.com/vsilab/dfu_ssl .

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multi-scale Attention Network for Diabetic Foot Ulcer Segmentation Using Self-supervised Learning

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Aravind Shrenivas Murali,
  • Marino Chuquilín,
  • Eung-Joo Lee

摘要

As part of the Diabetic Foot Ulcer Challenge (DFUC) 2024, we propose a novel self-supervised learning (SSL) approach for the segmentation of Diabetic Foot Ulcers (DFUs), leveraging a multi-scale feature extraction architecture combined with the DINOv2 model. This innovative approach is designed to address the variability and complexity in DFU images, which are often characterized by diverse shapes, sizes, and textures that pose significant challenges to traditional segmentation methods. Our method integrates an Attention U-NET as the segmentation head, enabling precise capture of critical ulcer features within DFU images, such as the boundaries and internal structures of ulcers. Using test datasets evaluated by the DFUC 2024 organizers, the model achieved a Dice coefficient of 0.5709 and a Jaccard index of 0.4793, demonstrating its instance segmentation capabilities on unseen data. The relevant code is released at: https://github.com/vsilab/dfu_ssl .