Development of a deep-learning model to detect free air on abdominal computed tomography for surgical decision support

Background: Free air (FA) in the abdominal cavity is a critical finding requiring prompt surgical intervention. We developed an AI-based segmentation model, Free Air-Net (FA-NET), to detect FA in abdominal computed tomography (CT) scans and further refined it with negative training to create FA-NET-NT, aiming to reduce false positives. Materials and methods: FA-NET-NT was developed using a retrospective dataset from a single institution ( n = 162). To evaluate its generalizability, the model was validated through both a temporal internal cohort ( n = 215) and an independent external cohort from a different hospital ( n = 237), which included various CT manufacturers and protocols. The model evaluation was threefold: (1) the Dice score coefficient, (2) image-wise, and (3) patient-wise sensitivity and specificity using representative CT segments (segments 4 through 8, out of 20 equally divided sections of the total axial series). If the model detected at least two images having FA among the representative images, the patient was regarded as having FA. Results: Both models achieved high Dice scores (0.87). FA-NET-NT improved specificity (96%) while maintaining high sensitivity (85%) in image-wise analysis. In patient-wise analysis, FA-NET-NT achieved 95–96% sensitivity for ulcer perforation and 82–92% specificity for non-FA conditions (cholecystitis, pancreatitis, and appendicitis). Specificity for ileus remained moderate (62%). In the external validation, the model demonstrated a patient-wise sensitivity of 95% for ulcer perforation. High specificity was maintained against differential diagnoses, including appendicitis (88%), pancreatitis (88%), cholecystitis (82%), and ileus (80%). Most false-positive findings were attributable to physiological bowel gas mimicking FA. Conclusion: FA-NET-NT is a robust decision-support tool for detecting FA, with its generalizability confirmed through multi-institutional validation. To provide definitive evidence of its clinical superiority, further prospective multi-center trials are necessary in emergency settings.