Federated attention contrastive learning models for prostate cancer diagnosis and Gleason grading

Artificial intelligence (AI) shows great promise in revolutionizing medical imaging, improving diagnosis, and refining treatment methods. However, the training of AI models relies on extensive multi-center datasets, presenting a potential challenge due to concerns about data privacy protection. Federated learning offers a solution by enabling a collaborative model across multiple centers without sharing raw data. In this study, we present a Federated Attention Contrastive Learning (FACL) framework designed to address challenges associated with large-scale pathological images and data heterogeneity. FACL improves model generalization by maximizing attention consistency between the local client and the server model. To enhance privacy and validate robustness, we incorporate differential privacy by introducing noise during parameter transfer. We assess the effectiveness of FACL in cancer diagnosis and Gleason grading tasks using 19,461 whole slide images of prostate cancer sourced from multiple centers. In the diagnosis task, FACL achieves an area under the curve (AUC) of 0.9718, outperforming seven centers whose average AUC is 0.9499 when categories are relatively balanced. In the Gleason grading task, FACL attained a Kappa score of 0.8463, surpassing the average Kappa score of 0.7379 from six centers. In conclusion, FACL offers a robust, accurate, and cost-effective AI training model for prostate cancer pathology while maintaining effective data safeguards.