Federated NephroNet: Privacy-Preserving Multi-Institutional Kidney CT Diagnosis Using Patient-Disjoint Decentralized Benchmarking and Localized Temperature Scaling Calibration

• Abilly Elly

  Texas University

  Author

Computed tomography (CT) is essential for diagnosing kidney pathologies, yet deep learning models for renal CT classification face critical barriers to clinical translation: centralized training requires sharing sensitive patient data across institutions, slice-level leakage from non-patient-disjoint splits inflates reported accuracy, and poorly calibrated models produce overconfident predictions that undermine clinical trust. This study introduces Federated NephroNet, a privacy-preserving framework that integrates federated learning with the patient-disjoint NephroNet benchmark and localized temperature scaling calibration. The methodology combines a compact depthwise-separable CNN (approximately 1.46M parameters) with FedAvg aggregation across simulated institutional clients, each performing local temperature scaling to calibrate predictions without sharing raw data. Evaluated on a multi-center kidney CT dataset (12,446 images; Normal, Cyst, Tumor, Stone), the federated model achieved 89.4% accuracy (95% CI: 88.7–90.1%) with macro-AUC of 0.954 and Expected Calibration Error of 0.031 after localized temperature scaling. While centralized NephroNet attained higher discrimination (0.9997 accuracy), the federated framework demonstrated substantially improved privacy guarantees and calibration robustness across heterogeneous client distributions. These findings establish that privacy-preserving multi-institutional kidney CT diagnosis is feasible with clinically acceptable calibration, providing a replicable architecture for decentralized medical AI deployment under regulatory constraints.

• pdf

Copyright (c) 2026 Abilly Elly (Author)

This work is licensed under a Creative Commons Attribution 4.0 International License.