Scholarlink

Purpose <p>To propose inter-disease out-of-domain generalization (OODG) across retinal diseases for microaneurysm (MA) segmentation using a deep-learning model trained and validated on diabetic retinopathy (DR) and qualitatively evaluated on leukemic retinopathy (LR).</p> Methods <p>A U-Net based segmentation model was trained using the IDRiD dataset, which comprises 81 DR images, using only MA annotations. The images were split into patches, and a statistical filtering step was applied to retain only structurally homogeneous patches. The study was organized in two phases: in Phase I, the U-Net was trained and evaluated using DR patches; in Phase II, the model was tested directly, without retraining, on LR images. Finally, MA segmentations were subjected to qualitative assessment by clinical specialists.</p> Results <p>The proposed U-Net achieved an IoU of 0.842, a Dice score of 0.914, an accuracy of 0.998, and a validation loss of 0.120.</p> Conclusion <p>The results suggest that knowledge learned from DR can generalize effectively to a related clinical context such as LR, opening the possibility of reusing models in diseases with similar structures and lesion patterns.</p>

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

Microaneurysm segmentation under out-of-domain generalization: from diabetic retinopathy to leukemic retinopathy

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年5月16日

查看全文
  • Edgar Platas-Campero,
  • Raquel Díaz-Hernández,
  • Anabel Sánchez-Sánchez,
  • Saúl Zapotecas-Martínez,
  • Nohemi Sánchez-Medel,
  • Sashwanthi Mohan,
  • Jamyl Habib-Castillo,
  • Leopoldo Altamirano-Robles

摘要

Purpose

To propose inter-disease out-of-domain generalization (OODG) across retinal diseases for microaneurysm (MA) segmentation using a deep-learning model trained and validated on diabetic retinopathy (DR) and qualitatively evaluated on leukemic retinopathy (LR).

Methods

A U-Net based segmentation model was trained using the IDRiD dataset, which comprises 81 DR images, using only MA annotations. The images were split into patches, and a statistical filtering step was applied to retain only structurally homogeneous patches. The study was organized in two phases: in Phase I, the U-Net was trained and evaluated using DR patches; in Phase II, the model was tested directly, without retraining, on LR images. Finally, MA segmentations were subjected to qualitative assessment by clinical specialists.

Results

The proposed U-Net achieved an IoU of 0.842, a Dice score of 0.914, an accuracy of 0.998, and a validation loss of 0.120.

Conclusion

The results suggest that knowledge learned from DR can generalize effectively to a related clinical context such as LR, opening the possibility of reusing models in diseases with similar structures and lesion patterns.