Scholarlink

<p>Diabetic Retinopathy (DR) causes permanent vision impairment among diabetic individuals globally. However, existing deep learning models for automated DR screening face challenges such as overfitting, limited generalizability, high computational requirements, and insufficient clinical interpretability. To address these problems, we introduce WaveMem-SHAPNet, an interpretable deep learning framework that incorporates advanced preprocessing, memory-aware optimization, and explainable AI components. Our two-step method improves fundus images through Adaptive Wavelet Denoising Transform (AWDT), median filtering, and Contrast Limited Adaptive Histogram Equalization (CLAHE). Subsequently, we utilize ResNet-50 integrated with Memory-Aware Synapse (MAS) regularization to maintain essential features while avoiding catastrophic forgetting. The framework combines Synaptic Adaptive Synapse Perturbation (SASP) to improve generalization and combines Swish-activated Grad-CAM with SHAP explanations to provide clinically interpretable predictions. We used three benchmark datasets to train and validate WaveMem-SHAPNet: APTOS (3,662 images), EyePACS (35,126 images), and IDRiD (516 images). The model outperformed current methods by 1–10%, achieving 98.74% accuracy, 97.85% precision, and 98.93% recall across all DR severity levels. Grad-CAM visualizations accurately localized hemorrhagic regions and exudates, while SHAP values evaluated feature contributions, thereby enabling transparent diagnostic decision-making.</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

WaveMem-SHAPNet: A Transparent Deep Learning Approach to Early Diagnosis of Diabetic Retinopathy

  • 期刊论文

  • 发表日期：

    2026年1月7日

查看全文
  • M. Kirubakaran,
  • V. Vijayarajan

摘要

Diabetic Retinopathy (DR) causes permanent vision impairment among diabetic individuals globally. However, existing deep learning models for automated DR screening face challenges such as overfitting, limited generalizability, high computational requirements, and insufficient clinical interpretability. To address these problems, we introduce WaveMem-SHAPNet, an interpretable deep learning framework that incorporates advanced preprocessing, memory-aware optimization, and explainable AI components. Our two-step method improves fundus images through Adaptive Wavelet Denoising Transform (AWDT), median filtering, and Contrast Limited Adaptive Histogram Equalization (CLAHE). Subsequently, we utilize ResNet-50 integrated with Memory-Aware Synapse (MAS) regularization to maintain essential features while avoiding catastrophic forgetting. The framework combines Synaptic Adaptive Synapse Perturbation (SASP) to improve generalization and combines Swish-activated Grad-CAM with SHAP explanations to provide clinically interpretable predictions. We used three benchmark datasets to train and validate WaveMem-SHAPNet: APTOS (3,662 images), EyePACS (35,126 images), and IDRiD (516 images). The model outperformed current methods by 1–10%, achieving 98.74% accuracy, 97.85% precision, and 98.93% recall across all DR severity levels. Grad-CAM visualizations accurately localized hemorrhagic regions and exudates, while SHAP values evaluated feature contributions, thereby enabling transparent diagnostic decision-making.