Scholarlink

<p>Diabetic retinopathy is a significant cause of vision loss; thus, accurately detecting exudates is essential for early diagnosis. However, exudates often resemble other retinal features such as the optic disc and haemorrhages, making automated segmentation challenging. Therefore, this study proposes a Hue-Saturation-Value (HSV) colour-based segmentation framework with preprocessing steps including Contrast Limited Adaptive Histogram Equalisation (CLAHE)-based contrast enhancement and optic disc masking to avoid false positives. Yellow exudates were segmented using optimised hue saturation thresholds. The approach was validated on the IDRID (Indian Diabetic Retinopathy Image Dataset) dataset, with evaluation based on pixel-level accuracy metrics. The method achieved 93.5% accuracy, 90.3% sensitivity, 93.3% specificity, 84.7% precision, 87.4% Dice score, and 77.6% Intersection over Union (IoU), demonstrating effective exudate localisation while reducing optic disc-related errors. The proposed framework provides reliable exudate segmentation and can serve as a preprocessing step for deep learning–based DR classification.</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

Exudate Detection and Segmentation Using a Colour and Illumination-Invariant HSV Model for Diabetic Retinopathy Classification

  • 期刊论文

  • 发表日期：

    2026年4月18日

查看全文
  • Jyoti Sawant,
  • Amol D. Vibhute

摘要

Diabetic retinopathy is a significant cause of vision loss; thus, accurately detecting exudates is essential for early diagnosis. However, exudates often resemble other retinal features such as the optic disc and haemorrhages, making automated segmentation challenging. Therefore, this study proposes a Hue-Saturation-Value (HSV) colour-based segmentation framework with preprocessing steps including Contrast Limited Adaptive Histogram Equalisation (CLAHE)-based contrast enhancement and optic disc masking to avoid false positives. Yellow exudates were segmented using optimised hue saturation thresholds. The approach was validated on the IDRID (Indian Diabetic Retinopathy Image Dataset) dataset, with evaluation based on pixel-level accuracy metrics. The method achieved 93.5% accuracy, 90.3% sensitivity, 93.3% specificity, 84.7% precision, 87.4% Dice score, and 77.6% Intersection over Union (IoU), demonstrating effective exudate localisation while reducing optic disc-related errors. The proposed framework provides reliable exudate segmentation and can serve as a preprocessing step for deep learning–based DR classification.