Multi-Window Deep Learning Detects Six Acute Abdominal Emergencies on CT

A multi-window deep learning system detected six acute abdominal emergencies on CT with macro AUROC 0.941 internally and 0.879 externally, but external F1 was 0.545 (0.648 after recalibration); the retrospective study requires prospective multisite validation.

• Retrospective study: 1274 patients for training/internal validation, external Stanford Merlin cohort of 280 patients.
• Nine-region abdominal localization accuracy 99.5% among detected cases, 90.9% including missed detections.
• Limitation: Reduced external F1 and need for site-specific calibration; prospective validation required before clinical deployment.

RadPigeon summaries are original and for information only. They are not clinical advice.