Abstract
Delays or misdiagnoses in detecting pneumoperitoneum can significantly increase mortality and morbidity. We developed and validated a deep learning model designed to identify pneumoperitoneum in computed tomography images. The model is trained on abdominal scans from Far Eastern Memorial Hospital (January 2012-December 2021) and evaluated using a simulated test set (14,039 scans) and a prospective test set (6351 scans) collected from the same center between December 2022 and May 2023. External validation included 480 scans from Cedars-Sinai Medical Center. Overall, the model achieves a sensitivity of 0.81-0.83 and a specificity of 0.97-0.99 across retrospective, prospective, and external validation; sensitivity improves to 0.92-0.98 when cases with a small amount of free air (total volume <10 ml) are excluded. These findings suggest that the model can deliver accurate and consistent predictions for pneumoperitoneum in computed tomography scans with segmented masks, potentially accelerating diagnostic and treatment workflows in emergency care.
Introduction
Pneumoperitoneum—the presence of "free air" within the abdominal cavity—is a surgical emergency typically signaling a perforated hollow organ. While Computed Tomography (CT) is the diagnostic gold standard, the surge in emergency department (ED) CT volume has led to significant interpretation delays. Furthermore, junior residents often face a steep learning curve in identifying subtle air pockets, with discrepancy rates between residents and attending radiologists reaching up to 30%.
To address these bottlenecks, we developed PACT-3D, a deep learning algorithm designed to automatically detect and segment pneumoperitoneum from 3D CT scans. PACT-3D acts as a high-speed triage tool, flagging critical cases and providing visual masks to guide clinical attention.
Methods and Architecture
PACT-3D utilizes a 3D U-Net convolutional neural network architecture. Unlike traditional 2D models that analyze individual slices, a 3D approach processes the entire volumetric stack. This allows the model to capture the complex 3D morphology of air pockets, helping it distinguish free air (which often collects under the diaphragm) from normal bowel gas.
Key technical features include:
- Hybrid Loss Function: A 50/50 combination of Dice loss (for segmentation) and Focal loss (to manage the extreme class imbalance where air occupies a tiny fraction of the scan).
- Dual Output: The system provides both a binary classification (positive/negative) and a segmented mask for localization.
Study Design and Datasets
The study leveraged a massive dataset of over 140,000 abdominal CT scans to ensure robustness and generalizability. The development cohort was sourced from Far Eastern Memorial Hospital (FEMH), while external validation was conducted using data from Cedars-Sinai Medical Center (USA) to test performance across different patient populations and scanner vendors (GE, Siemens, Philips, Toshiba).
Key Results
PACT-3D demonstrated high diagnostic accuracy across all test environments:
- Specificity: Remained consistently high at 0.97–0.99, minimizing "alarm fatigue" in busy clinical settings.
- Sensitivity: Ranged from 0.81–0.83 overall. However, for clinically significant air volumes (>10 ml), sensitivity improved dramatically to 0.92–0.98.
- Clinical Relevance: AI-positive cases were significantly more likely to require urgent surgery (75–85%) compared to false negatives, suggesting the model is naturally tuned to the most dangerous presentations.
- Segmentation Accuracy: The model achieved a Dice score of 0.81, indicating high overlap with expert manual annotations.
Discussion and Limitations
While PACT-3D excels at identifying clinically significant pneumoperitoneum, it still faces challenges with very small air bubbles (<1 ml) or air located in the retroperitoneal space. Occasionally, mimics such as skin emphysema or abscesses can trigger false positives.
Future work will focus on integrating this tool into real-time clinical workflows to measure its direct impact on patient outcomes, such as reducing the time from scan to surgical intervention. Currently, PACT-3D stands as a robust "second set of eyes," capable of providing consistent, accurate predictions to support emergency medical teams.
Summary updated