Scholarlink

The structure of the retinal vasculature can indicate various health issues. Quantitatively measuring changes in retinal arteries and veins offers significant potential for disease prevention and management. We propose VAVnets: three variants of a deep-learning network to generate Vessels, Arteries and Veins binary segmentations. Training is conducted in a few-shot and cross-dataset manner using images from four different fundus datasets, collectively comprising 137 images: DRIVE, DUMO, HRF, and LESAV. Training occurs with 3-fold cross-validation. Each fold employs a cross-dataset-build of 40 images (10 from each dataset), with testing on the remaining 97 images. In this article, we discuss our experiments involving architectural choices, transfer-learning, and data augmentation. We assess performances using the dice score as we aim to achieve the best possible pixel-wise segmentations. Our dice scores for each dataset are, for vessels: 0.81, 0.83, 0.81, 0.86; for veins: 0.78, 0.81, 0.78, 0.79; and for arteries: 0.73, 0.78, 0.74, 0.77. To the best of our knowledge, VAVnets demonstrate superior performances compared to existing few-shot methods across these datasets.

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

VAVnets: Retinal Vasculature Segmentation in Few-Shot Scenarios

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Idris Dulau,
  • Benoit Recur,
  • Catherine Helmer,
  • Cecile Delcourt,
  • Marie Beurton-Aimar

摘要

The structure of the retinal vasculature can indicate various health issues. Quantitatively measuring changes in retinal arteries and veins offers significant potential for disease prevention and management. We propose VAVnets: three variants of a deep-learning network to generate Vessels, Arteries and Veins binary segmentations. Training is conducted in a few-shot and cross-dataset manner using images from four different fundus datasets, collectively comprising 137 images: DRIVE, DUMO, HRF, and LESAV. Training occurs with 3-fold cross-validation. Each fold employs a cross-dataset-build of 40 images (10 from each dataset), with testing on the remaining 97 images. In this article, we discuss our experiments involving architectural choices, transfer-learning, and data augmentation. We assess performances using the dice score as we aim to achieve the best possible pixel-wise segmentations. Our dice scores for each dataset are, for vessels: 0.81, 0.83, 0.81, 0.86; for veins: 0.78, 0.81, 0.78, 0.79; and for arteries: 0.73, 0.78, 0.74, 0.77. To the best of our knowledge, VAVnets demonstrate superior performances compared to existing few-shot methods across these datasets.