Automated Structured Pre-Report Generation for Nephrolithiasis on Non-Contrast CT: Development and Report-Level Evaluation of a Deep Learning Framework.

This study aims to develop and evaluate a fully automated deep learning framework for generating structured preliminary radiology reports for nephrolithiasis assessment on non-contrast computed tomography (NCCT). This retrospective single-center study included 195 patients with radiologically confirmed nephrolithiasis on NCCT (390 kidneys). Kidneys were automatically localized and analyzed without manual image selection. Stone presence and size class were predicted using a hybrid CNN-Transformer with multiple-instance learning and compared with a ResNet-18 encoder, whereas intrarenal zone was derived from detector position. Structured pre-reports were generated using a deterministic template. Performance was evaluated using patient-level stratified fivefold cross-validation across five seeds and task-specific and report-level metrics. Kidney-level stone-presence AUC was 0.980 (95% CI, 0.975-0.986). Stone-size classification accuracy was 0.752 (95% CI, 0.725-0.780) with a quadratic weighted kappa of 0.726 (95% CI, 0.692-0.760), while intrarenal zone prediction macro AUC was 0.777 (95% CI, 0.751-0.803). Field-level agreement was 0.950 for stone presence, 0.876 for laterality, 0.737 for size class, and 0.728 for zone. Composite triplet-level F1 was 0.533, with hallucination and omission rates of 0.458 and 0.476, respectively. No primary metric showed a statistically supported difference between encoders; the stone-presence AUC difference was 0.002 (corrected 95% CI, - 0.015 to 0.019). Automated structured pre-report generation for nephrolithiasis without manual image selection was feasible. However, strong field-level performance did not translate into comparable report-level performance. No measurable incremental benefit of the hybrid CNN-Transformer over ResNet-18 was demonstrated in this cohort. External validation and prospective clinical assessment are required before clinical implementation.