Anatomically Localized Detection of Six Acute Abdominal Emergencies on CT Using Multi-window Deep Learning:...

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Acute abdominal emergencies require timely diagnosis for effective treatment.
• Deep learning models offer potential for automating the detection of these conditions on CT scans.

Purpose of the Study:

• To develop and validate a deep learning system for classifying and localizing six acute abdominal emergencies using CT.
• To assess the system's performance using multi-window Hounsfield Unit (HU) encoding and anatomical localization.

Main Methods:

• A YOLOv11-Large model was trained on a large national teleradiology dataset (1274 patients) with multi-window HU encoding.
• Localization was evaluated using a nine-region abdominal grid, and specificity was tested in a target-negative cohort.
• External validation was performed on the Stanford Merlin cohort (280 patients).

Main Results:

• Internal testing achieved a macro AUROC of 0.941 and macro F1 of 76.1%.
• Nine-region localization accuracy was high (99.5% for detected cases).
• External validation on the Stanford Merlin cohort showed a macro AUROC of 0.879, with moderate-to-high discrimination for all six classes.

Conclusions:

• The developed deep learning model demonstrates strong performance in detecting acute abdominal emergencies on CT.
• While promising, further prospective, multi-site validation and site-specific calibration are necessary before clinical implementation.
• Anatomical localization using the nine-region grid provides clinically relevant interpretability.