<p>The application of artificial intelligence (AI) in medical imaging has resulted in major advancements in patient outcomes, individualized treatment strategies, and diagnostic accuracy. However, the sensitive nature of medical data and strict privacy regulations pose challenges for data sharing across institutions. Federated Learning (FL) offers a solution by enabling collaborative model training without exchanging raw data, thereby preserving privacy. Despite these benefits, FL faces substantial challenges when dealing with heterogeneous data distributions commonly found in real-world medical scenarios, which can degrade model performance. To address this issue, this work introduces a novel FL framework that integrates Knowledge Distillation (KD) to address the challenges of data heterogeneity in medical imaging. Unlike traditional FL methods, our approach performs knowledge distillation at the feature level, prior to classification, enabling more effective knowledge transfer between the server and clients. Extensive experiments on three benchmark datasets–<i>TissueMNIST</i>, <i>PathMNIST</i>, and <i>BloodMNIST</i>–under two non-IID scenarios (Label Skew and Dirichlet distribution with <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="13748_2025_384_Article_IEq1.gif" Format="GIF" Height="14" Rendition="HTML" Resolution="72" Type="Linedraw" Width="55" /> </InlineMediaObject> <EquationSource Format="TEX">\(\alpha = 0.5\)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>α</mi> <mo>=</mo> <mn>0.5</mn> </mrow> </math></EquationSource> </InlineEquation>) demonstrate the effectiveness of the proposed method. Our FL-KD framework achieves up to <b>5.1%</b> higher accuracy than FedAvg and <b>3.5%</b> higher than FedProx, with the highest improvement observed on <i>TissueMNIST</i> (from <b>88.81%</b> to <b>93.33%</b> under Dirichlet). These results establish our method as a robust and privacy-preserving solution for real-world federated medical applications, offering improved generalization and stability across diverse and imbalanced data distributions.</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

Enhanced federated learning framework for handling heterogeneity in medical image data

  • Manjunath Naganna,
  • Guru Ramachandra Nayaka

摘要

The application of artificial intelligence (AI) in medical imaging has resulted in major advancements in patient outcomes, individualized treatment strategies, and diagnostic accuracy. However, the sensitive nature of medical data and strict privacy regulations pose challenges for data sharing across institutions. Federated Learning (FL) offers a solution by enabling collaborative model training without exchanging raw data, thereby preserving privacy. Despite these benefits, FL faces substantial challenges when dealing with heterogeneous data distributions commonly found in real-world medical scenarios, which can degrade model performance. To address this issue, this work introduces a novel FL framework that integrates Knowledge Distillation (KD) to address the challenges of data heterogeneity in medical imaging. Unlike traditional FL methods, our approach performs knowledge distillation at the feature level, prior to classification, enabling more effective knowledge transfer between the server and clients. Extensive experiments on three benchmark datasets–TissueMNIST, PathMNIST, and BloodMNIST–under two non-IID scenarios (Label Skew and Dirichlet distribution with \(\alpha = 0.5\) α = 0.5 ) demonstrate the effectiveness of the proposed method. Our FL-KD framework achieves up to 5.1% higher accuracy than FedAvg and 3.5% higher than FedProx, with the highest improvement observed on TissueMNIST (from 88.81% to 93.33% under Dirichlet). These results establish our method as a robust and privacy-preserving solution for real-world federated medical applications, offering improved generalization and stability across diverse and imbalanced data distributions.