Scholarlink

<p>Diabetic retinopathy (DR) is a diabetes complication that damages blood vessels in the retina, potentially leading to vision loss if left undiagnosed. Early detection via eye exams is crucial, but faster and automated analysis methods are essential in areas lacking expert ophthalmologists. However, limited patient data and privacy concerns prevent medical institutions from building robust models for automated DR diagnosis. Federated Learning (FL) addresses this by enabling collaborative training on decentralized datasets from multiple institutions without compromising patient privacy. This study investigates Swin Transformer-based FL frameworks for DR classification, using Federated Averaging and Federated Proximal as two aggregation strategies. To address data heterogeneity, we used three open-sourced datasets and established them as three participating clients in our framework. We have also created our private dataset of Bangladeshi patients by collecting data from Bangladesh Eye Hospital Ltd. (BEH), Bangladesh, and used it as a separate client for generalizability assessment of the trained server models. In our experiment, the FL aggregation strategies using the Transformer model are evaluated based on the performance of these frameworks in classifying the stages of DR using both the open-sourced and unseen patient data. Our Swin-FedAvg approach shows remarkable results compared to Swin-FedProx when evaluated on both the open-sourced datasets and the private BEH data. Furthermore, we present a comparison of the Swin Transformer-based frameworks with the existing frameworks based on these datasets and demonstrate the robustness of our proposed method.</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

A Swin Transformer-Based Federated Learning Approach for Classification of Diabetic Retinopathy

  • 期刊论文

  • 发表日期：

    2025年12月15日

查看全文
  • Mohammad Akbar Bin Shah,
  • Tanvir Azhar,
  • Asaduzzaman,
  • Mahfuzulhoq Chowdhury

摘要

Diabetic retinopathy (DR) is a diabetes complication that damages blood vessels in the retina, potentially leading to vision loss if left undiagnosed. Early detection via eye exams is crucial, but faster and automated analysis methods are essential in areas lacking expert ophthalmologists. However, limited patient data and privacy concerns prevent medical institutions from building robust models for automated DR diagnosis. Federated Learning (FL) addresses this by enabling collaborative training on decentralized datasets from multiple institutions without compromising patient privacy. This study investigates Swin Transformer-based FL frameworks for DR classification, using Federated Averaging and Federated Proximal as two aggregation strategies. To address data heterogeneity, we used three open-sourced datasets and established them as three participating clients in our framework. We have also created our private dataset of Bangladeshi patients by collecting data from Bangladesh Eye Hospital Ltd. (BEH), Bangladesh, and used it as a separate client for generalizability assessment of the trained server models. In our experiment, the FL aggregation strategies using the Transformer model are evaluated based on the performance of these frameworks in classifying the stages of DR using both the open-sourced and unseen patient data. Our Swin-FedAvg approach shows remarkable results compared to Swin-FedProx when evaluated on both the open-sourced datasets and the private BEH data. Furthermore, we present a comparison of the Swin Transformer-based frameworks with the existing frameworks based on these datasets and demonstrate the robustness of our proposed method.