Deep learning-based detection of pneumoperitoneum on computed tomography: performance and clinical feasibility.
Authors
Affiliations (5)
Affiliations (5)
- Division of Colon and Rectal surgery, Department of Surgery, Korea University Anam Hospital, Seoul, Korea, Republic of.
- Department of Radiology and Advanced Medical Imaging Institute, Korea University Anam Hospital, Seoul, Korea, Republic of.
- Department of Convergence Medicine, University of Ulsan College of Medicine, Asan Medical Center, Seoul, Korea, Republic of. [email protected].
- Inha University College of Medicine and Inha University Hospital, Incheon, Korea. [email protected].
- Division of Colon and Rectal surgery, Department of Surgery, Korea University Anam Hospital, Seoul, Korea, Republic of. [email protected].
Abstract
Early detection of pneumoperitoneum on computed tomography (CT) is critical because delayed diagnosis of gastrointestinal perforation is associated with increased morbidity and mortality. We evaluated the diagnostic performance and clinical applicability of a deep learning model for automated pneumoperitoneum detection on CT. In this retrospective diagnostic accuracy study, CT scans for 263 patients with surgically confirmed gastrointestinal perforation and 78 normal controls from a single tertiary center (January 2010-September 2024) were included. Positive CT scans were divided into training, validation, internal test, and temporally independent external test datasets, with 78 normal controls included in the external test dataset. Free air was manually segmented and verified by surgical and radiological experts. A segmentation-based deep learning model was developed using the no-new-Segment Anything Model (nnSAM) framework and compared with a conventional no-new-Net (nnUNet) model. Segmentation and classification performance were assessed on internal and temporally independent external test datasets. In the internal test dataset (n = 43 positive cases), the nnSAM model achieved a higher detection rate (97.7% vs. 90.7%) and significantly higher mean Dice similarity coefficient (0.739 ± 0.266 vs. 0.684 ± 0.283; p = 0.045) than the nnUNet one. In the external test dataset, nnSAM also showed a higher detection rate in positive cases (97.9% vs. 85.4%) and a significantly higher mean Dice similarity coefficient (0.863 ± 0.189 vs. 0.746 ± 0.208; p = 0.003). The performance advantage of nnSAM was particularly evident in small-volume pneumoperitoneum cases. The nnSAM-based model showed improved accuracy and robustness for pneumoperitoneum detection on CT, supporting its potential as a clinically useful decision-support tool for emergency radiology workflows and triage of patients with suspected gastrointestinal perforation.