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

NephroNet is a new artificial intelligence system designed to classify kidney pathologies on CT images into four categories: normal, cysts, tumors, and stones. The system was developed and tested on a dataset of 12,446 images from multiple medical centers. The NephroNet model is compact with 1.46 million parameters and uses specialized CNN architecture with squeeze-and-excitation and SpatialGate technologies. When tested on 2,490 unseen images, it achieved 99.97% accuracy with macro-AUC of 0.9969 and Brier score of 0.0007. The research emphasizes the importance of proper probability calibration and avoiding information leakage at the slice level. However, the authors note that all data come from a single region, so further external and prospective validation is needed.