Objective:

To create a patient-disjoint benchmark for four-class renal CT classification while addressing the challenges of slice-level leakage and inadequate probability calibration in deep learning assessments.

Approach:

Key Findings:

• NephroNet achieved an accuracy of 0.9997 (95% CI 0.9984–1.0000) in the validation set (N = 2,490).
• The model demonstrated a macro-AUC of 0.9969 (0.9953–0.9983), a Brier score of 0.0007, and an ECE of 0.0021.
• NephroNet outperformed budget-matched CNN and transformer baseline models.

Interpretation:

The study highlights the significance of patient-disjoint assessment and probability calibration in deep learning applications for renal CT classification.

Limitations:

• All findings are derived from a single-center geographic location (Dhaka, Bangladesh), which may limit generalizability.
• External and prospective validation is necessary to confirm the findings.

Conclusion:

The research establishes a controlled benchmark that can be utilized for future studies in renal CT classification.

Attribution Notice

This content is an AI-generated, fully rewritten summary based on a published scholarly article. It does not reproduce the original text and is not a substitute for the original publication. Readers are encouraged to consult the source for full context, data, and methodology.