Scholarlink

<p>U-Net exhibits limited capacity for delineating peripheral micro-vasculature and suppressing noise artefacts in fundus images; accordingly, we propose a hybrid framework that couples an augmented U-Net with a wavelet-domain discriminator to enhance retinal vessel segmentation. The upgraded U-Net functions as the generator, while both its predicted masks and their ground-truth counterparts are decomposed by discrete wavelet transform into high-frequency sub-bands and presented to the discriminator for adversarial learning. Through adversarial feedback, the discriminator drives progressive refinement of the generator outputs. Dense connectivity supplants conventional convolutional blocks within the encoder, simultaneously reducing parameters and enriching feature representation. Further performance gains are achieved by integrating squeeze-and-excitation attention and customised feature-fusion and attention-gate modules into the generator. Extensive experiments on CHASE_DB1, STARE, and DRIVE yield accuracies of 96.54%, 96.35%, and 96.42%; sensitivities of 80.16%, 77.34%, and 80.46%; specificities of 98.66%, 99.19%, and 98.72%; and AUCs of 98.57%, 98.14%, and 98.19%, respectively. These results demonstrate that the proposed architecture consistently surpasses baseline U-Net and Transformer-based Dual-Transformer models across all key metrics, underscoring its efficacy for retinal vessel segmentation.</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

U-Net Model Integrated with Wavelet Domain Discriminator for Retinal Vessel Segmentation

  • 期刊论文

  • 发表日期：

    2025年12月1日

查看全文
  • Haiyan Quan,
  • Zongxi Wang,
  • Leyi Zhang

摘要

U-Net exhibits limited capacity for delineating peripheral micro-vasculature and suppressing noise artefacts in fundus images; accordingly, we propose a hybrid framework that couples an augmented U-Net with a wavelet-domain discriminator to enhance retinal vessel segmentation. The upgraded U-Net functions as the generator, while both its predicted masks and their ground-truth counterparts are decomposed by discrete wavelet transform into high-frequency sub-bands and presented to the discriminator for adversarial learning. Through adversarial feedback, the discriminator drives progressive refinement of the generator outputs. Dense connectivity supplants conventional convolutional blocks within the encoder, simultaneously reducing parameters and enriching feature representation. Further performance gains are achieved by integrating squeeze-and-excitation attention and customised feature-fusion and attention-gate modules into the generator. Extensive experiments on CHASE_DB1, STARE, and DRIVE yield accuracies of 96.54%, 96.35%, and 96.42%; sensitivities of 80.16%, 77.34%, and 80.46%; specificities of 98.66%, 99.19%, and 98.72%; and AUCs of 98.57%, 98.14%, and 98.19%, respectively. These results demonstrate that the proposed architecture consistently surpasses baseline U-Net and Transformer-based Dual-Transformer models across all key metrics, underscoring its efficacy for retinal vessel segmentation.