Scholarlink

Inspired by the IFMBE Scientific Challenge Competition at NBC 2025, this work presents a deep learning approach for semantic wound segmentation employing a U-Net architecture. The approach focuses on segmenting three components within wound images: wound area, background, and scale marker. Preprocessing included grayscale conversion, resizing, and data augmentation on a dataset consisting of 371 training images and 80 validation images, each with associated 512 \(\times \) 512 pixel masks. On the validation set, the model gets a weighted F1 score of 0.8327 using an encoder-decoder structure with skip connections. Post-processing Lanczos interpolation guarantees that output masks satisfy the criteria of the challenge. The best model achieved a test set F1 score of 0.7924, as reported by the challenge organizers.

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

Deep Learning-Based Wound Segmentation Using a U-Net Architecture

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Martinus Boom

摘要

Inspired by the IFMBE Scientific Challenge Competition at NBC 2025, this work presents a deep learning approach for semantic wound segmentation employing a U-Net architecture. The approach focuses on segmenting three components within wound images: wound area, background, and scale marker. Preprocessing included grayscale conversion, resizing, and data augmentation on a dataset consisting of 371 training images and 80 validation images, each with associated 512 \(\times \) 512 pixel masks. On the validation set, the model gets a weighted F1 score of 0.8327 using an encoder-decoder structure with skip connections. Post-processing Lanczos interpolation guarantees that output masks satisfy the criteria of the challenge. The best model achieved a test set F1 score of 0.7924, as reported by the challenge organizers.