NephroNet: a calibration-aware, patient-disjoint benchmark for multiclass kidney CT classification with a compact depthwise-separable CNN - Report - MDSpire
Advertisement
NephroNet: a calibration-aware, patient-disjoint benchmark for multiclass kidney CT classification with a compact depthwise-separable CNN
Clinical Report: NephroNet: A Patient-Disjoint Benchmark for Kidney CT Classification
Overview
NephroNet, a depthwise-separable CNN, achieves an accuracy of 0.9997 in a validation set. This study addresses slice-level leakage and probability calibration in deep learning applications for renal imaging.
Background
Computed tomography (CT) is the primary imaging modality for renal pathology, yet challenges such as slice-level leakage and inadequate probability calibration hinder the effectiveness of deep learning models. This study introduces a patient-disjoint benchmark for classifying renal CT images into four categories: Normal, Cyst, Tumor, and Stone, using a multicenter dataset.
Data Highlights
Metric
Value
95% CI
Accuracy
0.9997
0.9984–1.0000
Macro-AUC
0.9969
0.9953–0.9983
Brier Score
0.0007
N/A
ECE
0.0021
N/A
Key Findings
NephroNet achieved an accuracy of 0.9997 in the validation set.
Utilized a multicenter dataset of 12,446 images for training and validation.
Implemented a patient-disjoint hold-out methodology to prevent slice-level leakage.
Demonstrated performance compared to budget-matched CNN and transformer baseline models.
Emphasized the importance of probability calibration.
Clinical Implications
The findings emphasize the necessity for external validation and calibration in diverse clinical settings.
Conclusion
Further validation is required to confirm the applicability of NephroNet in broader clinical contexts.
