不共享数据，也能联合训练，UCL团队用联邦学习重塑血液形态学检查

Core Insights - A research team from University College London (UCL) has developed a federated learning framework for white blood cell morphology analysis, enabling collaborative training without data exchange among institutions, thus ensuring data privacy while learning robust and domain-invariant feature representations [1][2]. Group 1: Federated Learning Framework - The federated model utilizes blood smear data from multiple clinical sites, demonstrating superior cross-site performance and generalization capabilities compared to centralized training [1][2]. - The framework addresses the critical issue of data privacy in healthcare, allowing institutions to collaborate on model training without sharing sensitive medical data [2][20]. Group 2: Clinical Relevance and Data Heterogeneity - Blood morphology examination is vital for diagnosing blood diseases, but it is labor-intensive and heavily reliant on skilled professionals, particularly in low- and middle-income countries where expertise is scarce [1]. - The study employed blood smear datasets from two centers, ensuring coverage of various cell types and reflecting real-world clinical heterogeneity, which is crucial for testing the federated learning model's generalization ability [5][8]. Group 3: Model Architecture and Training - The research utilized two deep learning architectures: ResNet-34 and DINOv2-Small, with a unified training protocol involving five rounds of global communication and local training cycles [9][11]. - Four federated aggregation strategies were implemented: FedAvg, FedMedian, FedProx, and FedOpt, each with distinct characteristics and performance implications [12]. Group 4: Performance Evaluation - The federated learning framework showed significant performance improvements, with models achieving a balanced accuracy of 58% compared to 52% for models trained on single institution data, highlighting the advantages of collaborative training without data sharing [16]. - External validation on data from a clinical hospital in Barcelona demonstrated that federated methods outperformed centralized training in generalization capabilities, achieving a balanced accuracy of 67% versus 64% [17][19]. Group 5: Implications for Healthcare - Federated learning is positioned as a key solution to overcome the "data silo" problem in healthcare, enabling collaborative model training while maintaining data privacy and compliance with regulations [20][22]. - The approach is expected to facilitate the transition of AI applications in blood morphology analysis from single-institution settings to cross-regional, clinical-grade intelligent diagnostic services, supporting precision medicine and digital healthcare [22].