NephroNet: a calibration-aware, patient-disjoint benchmark for multiclass kidney CT classification with a compact...

CT is the primary modality for kidney pathology, but deep-learning evaluation is undermined by slice-level leakage and poor probability calibration.

A patient-disjoint, group-stratified hold-out benchmark is established for four-class kidney CT classification (Normal, Cyst, Tumor, Stone) on a 12,446-image multicenter cohort, and NephroNet - a compact (1.46 M-parameter) depthwise-separable CNN with squeeze-and-excitation and a light SpatialGate -b is proposed. A standardized pipeline (320 × 320 preprocessing; annealed MixUp/CutMix; class-weighted AdamW with warmup-cosine; EMA-only evaluation; TTA; post-hoc temperature scaling, T^* = 1.42) is reported across accuracy, per-class and micro/macro ROC-AUC, Brier score, and ECE with bootstrap CIs.

On the hold-out (N = 2,490), NephroNet attains accuracy 0.9997 (95% CI 0.9984-1.0000), macro-AUC 0.9969 (0.9953-0.9983), Brier 0.0007, ECE 0.0021, surpassing budget-matched CNN and transformer baselines.

Transparent splits, explicit capacity control, and calibration-aware reporting support reproducible comparison; all data are single-region, so external and prospective validation is required.