Diagnostic performance of an artificial intelligence algorithm for detecting pneumoperitoneum on abdominal CT scans

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Pneumoperitoneum detection on CT scans is crucial for emergency triage.
• Current methods may face challenges in detecting small volumes of free air.
• AI offers potential for automated and efficient analysis of abdominal CT scans.

Purpose of the Study:

• To evaluate the diagnostic performance of an AI algorithm for detecting, segmenting, and quantifying pneumoperitoneum on abdominal CT scans.
• To assess the AI model's accuracy, sensitivity, specificity, and volume agreement with reference standards.
• To investigate the AI model's capability in identifying varying volumes of free air.

Main Methods:

• A deep learning model was developed and validated using multi-center CT imaging data from 2072 patients.
• The dataset was divided into training and testing sets, with external validation on 214 emergency CT scans.
• Diagnostic reports served as the reference standard; performance metrics included AUC, sensitivity, specificity, PPV, and NPV. Quantitative agreement was assessed using ICC.

Main Results:

• The AI model achieved high diagnostic performance in the test set (AUC 0.97) and external validation cohort (sensitivity 84.3%, specificity 89.6%).
• Performance remained robust, with improved sensitivity (96%) after excluding minimal free gas cases (<1 mL).
• AI-derived volumes showed strong agreement with the reference standard (ICC 0.996).

Conclusions:

• The AI model demonstrates high diagnostic accuracy for pneumoperitoneum on abdominal CT scans.
• The AI shows promise in expediting emergency radiology workflows and aiding triage decisions.
• Further prospective studies are needed to confirm the clinical impact of this AI tool.