Scholarlink

<p>Retinal diseases such as diabetic retinopathy (DR) and age-related macular degeneration (AMD) are increasingly prevalent worldwide, and their subtle early manifestations may delay screening and diagnosis, leading to irreversible vision loss and underscoring the importance of timely retinal disease analysis with optical coherence tomography angiography (OCTA). However, existing predictive modeling studies on OCTA images have made limited use of advanced machine learning methods and are predominantly based on conventional single-view strategies, which may overlook subtle pathological cues. As a result, effectively leveraging complementary information from multiple views for disease classification remains challenging. To address this issue, we developed a Multi-view Attentive Fusion and Adversarial Reconstruction Network (MAAR-Net) on OCTA-500 for retinal disease classification using six OCT/OCTA projection views. Specifically, a compression-attention fusion module (CFM) was designed to adaptively emphasize projections with clearer layer-specific pathological evidence and assign view-specific fusion weights. Considering the influence of motion artifacts, low signal quality, and segmentation errors, we introduced a semantic consistency-guided representation alignment (SCRA) mechanism to preserve disease-discriminative microvascular patterns across views in the latent space. In addition, to improve structural recoverability of the fused representation, we employed a view-wise adversarial reconstruction (VAR) strategy to reduce the loss of clinically meaningful pathological information. Extensive experiments demonstrated that MAAR-Net outperformed several conventional single-view baseline models in retinal disease diagnosis.</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

Multi-view attentive fusion and adversarial reconstruction network for retinal disease classification using OCT and OCTA

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月7日

查看全文
  • Yifei Wang,
  • Xing Jin,
  • Penggang Wang,
  • Yehao Fan,
  • Yiyuan Fu,
  • Xiyin Wu

摘要

Retinal diseases such as diabetic retinopathy (DR) and age-related macular degeneration (AMD) are increasingly prevalent worldwide, and their subtle early manifestations may delay screening and diagnosis, leading to irreversible vision loss and underscoring the importance of timely retinal disease analysis with optical coherence tomography angiography (OCTA). However, existing predictive modeling studies on OCTA images have made limited use of advanced machine learning methods and are predominantly based on conventional single-view strategies, which may overlook subtle pathological cues. As a result, effectively leveraging complementary information from multiple views for disease classification remains challenging. To address this issue, we developed a Multi-view Attentive Fusion and Adversarial Reconstruction Network (MAAR-Net) on OCTA-500 for retinal disease classification using six OCT/OCTA projection views. Specifically, a compression-attention fusion module (CFM) was designed to adaptively emphasize projections with clearer layer-specific pathological evidence and assign view-specific fusion weights. Considering the influence of motion artifacts, low signal quality, and segmentation errors, we introduced a semantic consistency-guided representation alignment (SCRA) mechanism to preserve disease-discriminative microvascular patterns across views in the latent space. In addition, to improve structural recoverability of the fused representation, we employed a view-wise adversarial reconstruction (VAR) strategy to reduce the loss of clinically meaningful pathological information. Extensive experiments demonstrated that MAAR-Net outperformed several conventional single-view baseline models in retinal disease diagnosis.