Scholarlink

<p>The diabetic retinopathy (DR) is one of the key causes of vision loss and blindness which can be effectively treated only in case of its early detection. The present paper introduces a novel and effective ensemble deep learning-based model to automatically detect the stages of DR. The suggested framework combines two complementary base learners, an EfficientNet-B0 classifier (Model 1) and a RetinaNet-based classifier trained on focal loss with ResNet-50 as the backbone (Model 2) with a weighted sum aggregation rule that puts a larger weight on a model that has a better validation performance. The resulting ensemble correctly identifies five different stages of DR (No DR, Mild, Moderate, Severe, and Proliferative) on a class-balanced subset of the EyePACS and the APTOS 2019 datasets with a Quadratic Weighted Kappa (QWK) of 0.932. The comparative performance with recent deep-learning baselines suggests that the ensemble can effectively leverage the complementary advantage of the two architectures, thus enhancing the accuracy and clinical performance of automated DR detection. The system design is to facilitate bulk screening under a resource-limited healthcare setting.</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

Multi-Class Diabetic Retinopathy Screening Using Effective Ensemble Approach

  • 期刊论文

  • 发表日期：

    2026年6月1日

查看全文
  • Punam Sunil Raskar,
  • Vishal Naranje,
  • Swaroop Pillai,
  • Shital Ashok Pawar

摘要

The diabetic retinopathy (DR) is one of the key causes of vision loss and blindness which can be effectively treated only in case of its early detection. The present paper introduces a novel and effective ensemble deep learning-based model to automatically detect the stages of DR. The suggested framework combines two complementary base learners, an EfficientNet-B0 classifier (Model 1) and a RetinaNet-based classifier trained on focal loss with ResNet-50 as the backbone (Model 2) with a weighted sum aggregation rule that puts a larger weight on a model that has a better validation performance. The resulting ensemble correctly identifies five different stages of DR (No DR, Mild, Moderate, Severe, and Proliferative) on a class-balanced subset of the EyePACS and the APTOS 2019 datasets with a Quadratic Weighted Kappa (QWK) of 0.932. The comparative performance with recent deep-learning baselines suggests that the ensemble can effectively leverage the complementary advantage of the two architectures, thus enhancing the accuracy and clinical performance of automated DR detection. The system design is to facilitate bulk screening under a resource-limited healthcare setting.