Federated Learning With AutoAlbum for Kidney Disease Detection: A Privacy-Preserving Approach to Medical Image Analysis

The increasing prevalence of kidney-related diseases necessitates efficient and accurate automated diagnostic tools to overcome the limitations of time-consuming and subjective manual image analysis. This study introduces a federated learning framework for decentralized kidney disease detection from CT scans, using AutoAlbum, an automated data augmentation technique, to introduce realistic variations through a diverse set of 11 augmentations. AutoAlbum enhances data variability and simulates real-world conditions, improving model robustness. The original Kidney Dataset (KD) is augmented globally using AutoAlbum to create a balanced dataset (AKD) for pre-trained model selection. KD is also partitioned into distributed subsets (KD-D) to mimic decentralized data in federated learning. Furthermore, augmentations are applied to these distributed KD subsets, resulting in the augmented and distributed dataset (AD-D), designed to evaluate the federated model’s performance under Non-Identical and Independent Distribution (Non-IID) conditions. KD represents the original data, AKD the globally augmented data, KD-D the distributed original data, and AD-D the distributed and augmented data. The federated learning framework enables privacy-preserving collaborative training across KD-D and AD-D. Evaluations showed steady performance improvements on KD-D, reaching 0.9663 accuracy by round five. Notably, the model demonstrated remarkable robustness on the more challenging AD-D, achieving an exceptional accuracy of 0.9916 by the fifth round, highlighting the framework’s potential for handling data heterogeneity in real-world clinical settings. To support model interpretability and clinical relevance, Grad-CAM visualizations were employed to reveal class-specific activation regions, enhancing transparency in model decision-making.