Scholarlink

<p>Early detection of eye diseases such as diabetic retinopathy, glaucoma, and macular degeneration is critical for preventing vision loss and to ensure timely diagnosis. Centralized training of deep neural networks for medical image analysis is often constrained by data privacy regulations and institutional boundaries. Federated learning (FL) allows model training to be decentralized without exchange of raw data. This research is a structured comparison of pretrained convolutional neural networks (CNN) AlexNet, VGG16, ResNet50, and EfficientNet-B7 with a vision transformer (ViT) in centralized and federated training schemes. The effects of heterogeneous and non-identically distributed (non-IID) client data on architectural stability remains unaddressed in ophthalmic classification. Non-IID client distributions are simulated using Dirichlet-based partitioning to model realistic hospital-level label skew. This study focuses on examining the response of various architectures to decentralized optimization when the client distribution is non-uniform, focusing in particular on the stability of global models across training rounds. The sensitivity of global models to non-IID label skew and the federated aggregation rate at which global models converge to the correct solution are the main challenges addressed in this work. The results demonstrate substantial variability in architectural robustness under non-IID federated conditions, indicating that model depth and structural complexity significantly influence decentralized convergence. These are supplemented to the practical knowledge of federated learning in classification of ophthalmic images.</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

Architecture sensitivity analysis of CNN and vision transformer models for federated eye disease classification under non-IID settings

  • 期刊论文

  • 发表日期：

    2026年6月3日

查看全文
  • Manjil Nepal,
  • Krishna Siva Prasad Mudigonda,
  • Pandu Sowkuntla,
  • Ashu Abdul

摘要

Early detection of eye diseases such as diabetic retinopathy, glaucoma, and macular degeneration is critical for preventing vision loss and to ensure timely diagnosis. Centralized training of deep neural networks for medical image analysis is often constrained by data privacy regulations and institutional boundaries. Federated learning (FL) allows model training to be decentralized without exchange of raw data. This research is a structured comparison of pretrained convolutional neural networks (CNN) AlexNet, VGG16, ResNet50, and EfficientNet-B7 with a vision transformer (ViT) in centralized and federated training schemes. The effects of heterogeneous and non-identically distributed (non-IID) client data on architectural stability remains unaddressed in ophthalmic classification. Non-IID client distributions are simulated using Dirichlet-based partitioning to model realistic hospital-level label skew. This study focuses on examining the response of various architectures to decentralized optimization when the client distribution is non-uniform, focusing in particular on the stability of global models across training rounds. The sensitivity of global models to non-IID label skew and the federated aggregation rate at which global models converge to the correct solution are the main challenges addressed in this work. The results demonstrate substantial variability in architectural robustness under non-IID federated conditions, indicating that model depth and structural complexity significantly influence decentralized convergence. These are supplemented to the practical knowledge of federated learning in classification of ophthalmic images.