Scholarlink

Diabetic retinopathy (DR) is a top contributor to visual impairment, for which early diagnosis is essential to impact effectively. Advanced Deep Learning (DL) methods are impeded by low interpretability and unbalanced class handling, thereby impairing clinical feasibility. In this work, we present an innovative DL framework that combines Vision Transformers (ViT) with Convolutional Neural Networks (CNNs) to automate the initial screening procedure for DR. The goal is to improve the diagnostic efficiency and accuracy, thereby helping ophthalmologist in making faster and more reliable choices. The proposed model uses attention maps from ViT and integrates heterogeneous activation functions within CNN to enable effective extraction of features. The dataset featured in Aptos Kaggle competition was used for this study. The hybrid model was evaluated using accuracy, AUC-ROC, sensitivity, and specificity. The model performed remarkably well demonstrating that our approach enhances classification accuracy and interpretability, offering a robust solution for automated DR screening. It can be derived that the model can reliably identify and distinguish among severities of DR.

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

Enhancing Ocular Disease Diagnosis: An Attention-Guided CNN for Diabetic Retinopathy Detection Using Vision Transformer-Derived Channel Augmentation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • B. S. Bhargava,
  • Sathiya Narayanan

摘要

Diabetic retinopathy (DR) is a top contributor to visual impairment, for which early diagnosis is essential to impact effectively. Advanced Deep Learning (DL) methods are impeded by low interpretability and unbalanced class handling, thereby impairing clinical feasibility. In this work, we present an innovative DL framework that combines Vision Transformers (ViT) with Convolutional Neural Networks (CNNs) to automate the initial screening procedure for DR. The goal is to improve the diagnostic efficiency and accuracy, thereby helping ophthalmologist in making faster and more reliable choices. The proposed model uses attention maps from ViT and integrates heterogeneous activation functions within CNN to enable effective extraction of features. The dataset featured in Aptos Kaggle competition was used for this study. The hybrid model was evaluated using accuracy, AUC-ROC, sensitivity, and specificity. The model performed remarkably well demonstrating that our approach enhances classification accuracy and interpretability, offering a robust solution for automated DR screening. It can be derived that the model can reliably identify and distinguish among severities of DR.