Domain-Invariant and Calibration-Preserving Deep Learning for Multi-Center Kidney CT Classification: Evaluating NephroNet Architecture Across Heterogeneous Scanners and Contrast Phase Dynamics

• Abilly Elly

  Texas University

  Author

Computed tomography (CT) remains the primary imaging modality for renal pathology, yet deep learning models for kidney CT classification face persistent barriers to clinical translation, including slice-level data leakage, poor probability calibration, and limited generalizability across heterogeneous scanner platforms. This study presents a comprehensive evaluation of NephroNet, a compact depthwise-separable convolutional neural network with squeeze-and-excitation channel attention and spatial gating, within a rigorously controlled patient-disjoint benchmark for four-class kidney CT classification (Normal, Cyst, Tumor, Stone). Using a 12,446-image multi-center cohort from Dhaka, Bangladesh, we implemented group-stratified hold-out validation to prevent patient-series leakage, combined with a standardized training pipeline incorporating annealed MixUp/CutMix augmentation, class-weighted AdamW optimization, exponential moving average evaluation, test-time augmentation, and post-hoc temperature scaling (T* = 1.42). NephroNet achieved an accuracy of 0.9997 (95% CI: 0.9984–1.0000), macro-AUC of 0.9969 (95% CI: 0.9953–0.9983), Brier score of 0.0007, and expected calibration error of 0.0021, surpassing 30 capacity-matched CNN and Vision Transformer baselines. These findings demonstrate that parameter-efficient architectural design combined with calibration-aware evaluation protocols can achieve near-ceiling discrimination while maintaining probability reliability, establishing a reproducible benchmark for domain-invariant renal imaging AI.

• pdf

Copyright (c) 2026 Abilly Elly (Author)

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