Scholarlink

<p>Diabetic retinopathy (DR) is a severe microvascular complication caused by damage to tiny blood vessels in the retina, leading to vision impairment and in severe cases blindness. Analysing the growth and changes in retinal blood vessels is essential for understanding the progression of DR. Precise segmentation of these vessels is crucial for effective diagnosis and monitoring. The present article presents a robust blood vessel segmentation framework that improves upon existing methods by capturing both thick and thin vessel structures while ensuring robustness against contrast variations. The framework addresses a common gap in current deep learning models, which often focus on building complex architectures without explicitly leveraging the curvilinear nature of blood vessels. The proposed method combines existing relative intensity order transformation (RIOT) to effectively extract thick and thin vessels and additionally texture features are incorporated to enhance contrast invariance. By processing 16 pixels in both horizontal and vertical directions, the model captures the underlying curvilinear structure of vessels, ensuring that the segmentation is robust to variations in image contrast and vessel thickness. The proposed method integrates modified IterNet, a U-Net based network to iteratively refine vessel segmentation for improved accuracy and continuity. The model was validated on the CHASEDB1, DRIVE and STARE retinal image datasets. It achieved segmentation accuracies of 99% on CHASEDB1, 96.5% on DRIVE and 88% on STARE, outperforming several state-of-the-art models in terms of both accuracy and robustness<b>.</b> Additionally, the proposed approach demonstrated superior performance in key evaluation metrics: sensitivity (83.2%, 96.3%, 75.2%), specificity (97.9%, 99%, 92%), F1-score (82.3%, 93%, 63%) DRIVE, CHASEDB1 and STARE respectively. By capturing the unique curvilinear structure of blood vessels, the proposed model—incorporating modified IterNet with RIOT and texture based features—offers improved segmentation performance, addressing the limitations of prior approaches that rely primarily on deep learning architectures.</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

Retinal blood vessel detection using deep learning with texture features and intensity order transformation

  • 期刊论文

  • 发表日期：

    2025年10月15日

查看全文
  • Chandrakala Kuruba,
  • Michael Arock

摘要

Diabetic retinopathy (DR) is a severe microvascular complication caused by damage to tiny blood vessels in the retina, leading to vision impairment and in severe cases blindness. Analysing the growth and changes in retinal blood vessels is essential for understanding the progression of DR. Precise segmentation of these vessels is crucial for effective diagnosis and monitoring. The present article presents a robust blood vessel segmentation framework that improves upon existing methods by capturing both thick and thin vessel structures while ensuring robustness against contrast variations. The framework addresses a common gap in current deep learning models, which often focus on building complex architectures without explicitly leveraging the curvilinear nature of blood vessels. The proposed method combines existing relative intensity order transformation (RIOT) to effectively extract thick and thin vessels and additionally texture features are incorporated to enhance contrast invariance. By processing 16 pixels in both horizontal and vertical directions, the model captures the underlying curvilinear structure of vessels, ensuring that the segmentation is robust to variations in image contrast and vessel thickness. The proposed method integrates modified IterNet, a U-Net based network to iteratively refine vessel segmentation for improved accuracy and continuity. The model was validated on the CHASEDB1, DRIVE and STARE retinal image datasets. It achieved segmentation accuracies of 99% on CHASEDB1, 96.5% on DRIVE and 88% on STARE, outperforming several state-of-the-art models in terms of both accuracy and robustness. Additionally, the proposed approach demonstrated superior performance in key evaluation metrics: sensitivity (83.2%, 96.3%, 75.2%), specificity (97.9%, 99%, 92%), F1-score (82.3%, 93%, 63%) DRIVE, CHASEDB1 and STARE respectively. By capturing the unique curvilinear structure of blood vessels, the proposed model—incorporating modified IterNet with RIOT and texture based features—offers improved segmentation performance, addressing the limitations of prior approaches that rely primarily on deep learning architectures.