PACT-3D, a Deep Learning Algorithm for Pneumoperitoneum Detection in Abdominal CT Scans

Abstract

Pneumoperitoneum, necessitates surgical intervention in 85-90% of cases, relies heavily on CT scans for diagnosis. Delay or misdiagnosis in detecting pneumoperitoneum can significantly increase mortality and morbidity. Our study introduced PACT-3D, a deep learning model developed to identify pneumoperitoneum in CT images. In this single hospital study, we retrospectively reviewed abdominal CT scans from January 2012 to December 2021, excluded CT of image acquisition error and without reports to form the development dataset for training the model. We evaluated the PACT- 3D model using a simulated test set of 14,039 scans and a prospective test set of 6,351 scans, collected from December 2022 to May 2023 at the same center. PACT-3D achieved a sensitivity of 0.81 and a specificity of 0.99 in retrospective testing, with prospective validation yielding a sensitivity of 0.83 and a specificity of 0.99. Sensitivity improved to 0.95 and 0.98 when excluding cases with a small amount of free air (total volume < 10ml) in simulated and prospective test sets, respectively. By delivering accurate and consistent patient-level predictions and providing segmented masks, PACT- 3D holds significant potential for assisting rapid decision-making in emergency care, thereby potentially improving patient outcomes.