Federated learning with comparative learning-based dynamic parameter updating on glioma whole slide images

• Published in: Engineering applications of artificial intelligence Vol. 137; p. 109233
• Main Authors: Huang, Longjian, Shao, Lizhi, Bao, Meiling, Guo, Changsong, Shao, Zhuhong, Huang, Xiazi, Wang, Mingjing, Jiang, Xiaoming, Hu, Shengzhou
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.11.2024
• Subjects: Deep learning; Federated learning; Pathology; Privacy protect; Whole slide imaging; Deep learning; Pathology; Privacy protect; Federated learning; Whole slide imaging
• ISSN: 0952-1976
• DOI: 10.1016/j.engappai.2024.109233

The rapid advancements in artificial intelligence have profoundly impacted various societal domains, particularly in healthcare. In computational pathology, deep learning techniques have shown remarkable abilities in classifying, segmenting, and recognizing pathology images. However, acquiring large-scale, high-quality medical datasets has become challenging due to increased privacy concerns and data protection awareness among institutions and patients. We propose utilizing federated learning to address this data privacy issue in this study. Our research focuses on classifying glioma whole slide images. To enhance the privacy of sensitive data, we incorporate Laplace noise into the model parameters of each local client. This technique guarantees the protection of patients’ data while allowing collaborative learning. Moreover, we introduce a novel method called Federated Learning with Comparative Learning-based Dynamic Parameter Updating. We select a local model with the optimal performance before all local model parameters are aggregated into global model parameters. Other local models then learn to update parameters from this selected model. By incorporating the Comparative Learning-based Dynamic Parameter Updating, we enhance the learning effect and improve the overall model performance for classifying glioma data. To assess our proposed method, we perform assessments on two separate classification tasks. The results of our experiments show that our privacy-preserving federated learning framework effectively utilizes multi-center data while maintaining good privacy protection performance. Additionally, compared to the commonly used federated averaging baseline method, our approach significantly outperforms glioma data classification tasks. Our research offers a promising framework that achieves high classification accuracy and ensures the protection of sensitive medical data, thus showcasing its potential in advancing computational pathology research and practice. Our code is free at https://github.com/jiaxian-hlj/FL-Dpu.