FedStoneNet-Hybrid: A Privacy-Preserving Federated CNN–Transformer Framework for Kidney Stone Detection from CT Imaging

Abstract

One of the most frequent urological issues that calls for accurate, timely diagnosis and appropriate treatment is renal stone disease, and CT is very helpful in diagnosing renal stones. Methods of deep learning possess effective automated detection potential. However, most prevailing methods involve centralized training, raising privacy and multi-institutional collaboration concerns. In response to these challenges, this research presents FedStoneNet-Hybrid, a federated learning framework designed to protect the privacy of kidney stone detection on distributed CT data using a hybrid CNN and transformer. Under the proposed scheme, each institution will locally train its hybrid model, where CNN layers will capture a fine-grained spatial representation and transformer modules will capture the global contextual relationship of pollution data. The Federated Averaging algorithm aggregates only the model parameters at the central server rather than the raw medical data. Thus, safe and scalable collaborative learning can be enabled. As observed from the experimental results, the proposed model achieves good performance with an accuracy of 97.2%, sensitivity of 96.8%, specificity of 96.9%, F1 score of 0.96.85, and AUC of 0.991. The model shows solid calibration with a low Expected Calibration Error and Brier Score. Based on these findings, the suggested framework serves as an accurate, scalable, and privacy-preserving approach for distributed medical imaging.