FedSAF: A Federated Learning Framework for Enhanced Gastric Cancer Detection and Privacy Preservation

Gastric cancer is one of the most commonly diagnosed cancers and has a high mortality rate. Due to limited medical resources, developing machine learning models for image-based gastric cancer recognition provides an efficient solution for medical institutions. However, such models typically require large sample sizes, which can challenge patient privacy. Federated learning enables collaborative model training across institutions without sharing sensitive data. This paper addresses limited public data through a modified data processing method. We systematically evaluate multiple pre-trained computer vision models and select the most suitable architecture based on classification performance and computational efficiency for gastric cancer image analysis. We then introduce FedSAF, a federated learning method integrating split personalization, similarity-weighted aggregation, and tFIM-based straggler down-weighting within a two-stage server aggregation pipeline to improve performance under non-IID settings. FedSAF incorporates attention-based message passing and Fisher Information Matrix weighting while reducing computation and transmission costs through model splitting. On gastric cancer datasets, FedSAF improves best test accuracy from 0.9790 to 0.9843 on SEED and from 0.7834 to 0.8116 on BOT compared with FedAMP. Model splitting reduces transmitted parameters by 83.70% while maintaining high accuracy (0.9798). The framework demonstrates strong robustness across SEED, BOT, FashionMNIST, and CIFAR-10. The proposed aggregation strategy is applicable to privacy-sensitive distributed learning scenarios.