Scholarlink

<p>Diabetic Retinopathy (DR) remains a leading cause of vision loss globally, necessitating accurate and scalable diagnostic solutions. Existing Deep Learning (DL) models often underutilize lesion-specific cues that are critical for early DR grading, while detection based models require costly lesion annotations. To address these limitations, we propose DRCNN-Lesion Proxy, a hybrid architecture that integrates a ResNet34 based CNN backbone for extracting global image level features with a Lesion Proxy Module, which simulates lesion-inspired cues without explicit lesion bounding box annotations. These heterogeneous features are fused through a late fusion classification head to enable robust multiclass DR severity prediction. The model was trained on a composite dataset and rigorously evaluated across six publicly available benchmarks namely EyePACS, Messidor-2, APTOS 2019, DDR, DIARETDB1, and IDRiD. Experimental findings show that the proposed framework consistently outperforms baseline CNNs and recent hybrid methods, achieving up to 98.37% accuracy, 97.28% F1-score, and 98.14% AUC. Statistical significance testing confirmed that these improvements were not due to chance. Furthermore, Grad-CAM visualizations highlighted clinically relevant retinal regions, and a pilot validation with three ophthalmologists on 20 cases reported mean scores above 3.5 out of 5, confirming that the explanations were perceived as clinically meaningful and useful for grading. The proposed framework provides an annotation light solution with strong generalizability, diagnostic precision, and clinically validated interpretability, advancing the state of the art in automated DR screening and offering a practical pathway for real world deployment.</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

DRCNN-Lesion Proxy: a hybrid CNN with lesion-inspired feature simulation for diabetic retinopathy severity classification

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年10月30日

查看全文
  • Priyadharshini Sekar,
  • Kanaga S. Suba Raja,
  • Ramaswamy Krishnaraj

摘要

Diabetic Retinopathy (DR) remains a leading cause of vision loss globally, necessitating accurate and scalable diagnostic solutions. Existing Deep Learning (DL) models often underutilize lesion-specific cues that are critical for early DR grading, while detection based models require costly lesion annotations. To address these limitations, we propose DRCNN-Lesion Proxy, a hybrid architecture that integrates a ResNet34 based CNN backbone for extracting global image level features with a Lesion Proxy Module, which simulates lesion-inspired cues without explicit lesion bounding box annotations. These heterogeneous features are fused through a late fusion classification head to enable robust multiclass DR severity prediction. The model was trained on a composite dataset and rigorously evaluated across six publicly available benchmarks namely EyePACS, Messidor-2, APTOS 2019, DDR, DIARETDB1, and IDRiD. Experimental findings show that the proposed framework consistently outperforms baseline CNNs and recent hybrid methods, achieving up to 98.37% accuracy, 97.28% F1-score, and 98.14% AUC. Statistical significance testing confirmed that these improvements were not due to chance. Furthermore, Grad-CAM visualizations highlighted clinically relevant retinal regions, and a pilot validation with three ophthalmologists on 20 cases reported mean scores above 3.5 out of 5, confirming that the explanations were perceived as clinically meaningful and useful for grading. The proposed framework provides an annotation light solution with strong generalizability, diagnostic precision, and clinically validated interpretability, advancing the state of the art in automated DR screening and offering a practical pathway for real world deployment.