Objectives This study aims to evaluate the diagnostic performance of an artificial intelligence (AI) algorithm for detection, segmentation, and volumetric quantification of pneumoperitoneum on abdominal CT scans. Materials and methods We developed and validated a deep learning-based model for automated pneumoperitoneum detection on CT. Multi-center CT imaging series from 2072 patients were collected and randomly divided into training and testing sets at an approximate 7:3 ratio. The external validation set included 214 emergency CT scans collected between April 2022 and December 2024. Diagnostic reports served as the reference standard. Primary outcome included the area under the curve (AUC), sensitivity, specificity, accuracy, positive predictive value (PPV), and negative predictive value (NPV). Quantitative agreement between AI and reference volumes was assessed using the intraclass correlation coefficient (ICC). Results In the test set (n = 607), the model demonstrated excellent performance: sensitivity 91.4%, specificity 93.1%, and AUC 0.97 (95% CI: 0.95–0.99). In the external validation cohort (n = 214), the model maintained robust performance with sensitivity 84.3% (95% CI: 76.2–90.5%), specificity 89.6% (95% CI: 82.3–94.6%), accuracy 86.9% (81.6–91.2%), PPV 89.2% (95% CI: 81.8–94.3%), and NPV 84.8% (95% CI: 77.1–90.7%). After excluding cases with minimal free gas (<1 mL), the model’s sensitivity improved to 96%. AI-derived volumes showed strong agreement with the reference standard (ICC 0.996, 95% CI: 0.994–0.997). Conclusion The AI model attained high diagnostic accuracy for pneumoperitoneum on abdominal CT scans, promising to expedite emergency workflows. Key Points <Emphasis Type="BoldItalic">Question</Emphasis> Reliable AI detection of pneumoperitoneum, particularly for small-volume free air, on emergency CT remains an unmet need for rapid and accurate emergency triage. <Emphasis Type="BoldItalic">Findings</Emphasis> The AI model shows high sensitivity and specificity for clinically relevant pneumoperitoneum volumes, although trace-volume detection on CT scans remains challenging. <Emphasis Type="BoldItalic">Critical relevance statement</Emphasis> This study evaluates the diagnostic performance and volume-dependent variability of an AI model for pneumoperitoneum detection on CT scans, with the potential to aid emergency radiology workflow prioritization and decision support, though prospective studies are needed to confirm clinical impact. Graphical Abstract

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

期刊论文
出版模式：

开放获取
发表日期：

2026年7月18日

查看全文

Yuwan Hu,
Zhigang Sun,
Haoyu Li,
Shuya Gao,
Haoran Du,
Tongyin Zhang,
Sheng Xie,
Peng Wu,
Pengbo Jiang,
Dijia Wu,
Xuemeng Wu,
Hongliang Sun

摘要

Objectives

This study aims to evaluate the diagnostic performance of an artificial intelligence (AI) algorithm for detection, segmentation, and volumetric quantification of pneumoperitoneum on abdominal CT scans.

Materials and methods

We developed and validated a deep learning-based model for automated pneumoperitoneum detection on CT. Multi-center CT imaging series from 2072 patients were collected and randomly divided into training and testing sets at an approximate 7:3 ratio. The external validation set included 214 emergency CT scans collected between April 2022 and December 2024. Diagnostic reports served as the reference standard. Primary outcome included the area under the curve (AUC), sensitivity, specificity, accuracy, positive predictive value (PPV), and negative predictive value (NPV). Quantitative agreement between AI and reference volumes was assessed using the intraclass correlation coefficient (ICC).

Results

In the test set (n = 607), the model demonstrated excellent performance: sensitivity 91.4%, specificity 93.1%, and AUC 0.97 (95% CI: 0.95–0.99). In the external validation cohort (n = 214), the model maintained robust performance with sensitivity 84.3% (95% CI: 76.2–90.5%), specificity 89.6% (95% CI: 82.3–94.6%), accuracy 86.9% (81.6–91.2%), PPV 89.2% (95% CI: 81.8–94.3%), and NPV 84.8% (95% CI: 77.1–90.7%). After excluding cases with minimal free gas (<1 mL), the model’s sensitivity improved to 96%. AI-derived volumes showed strong agreement with the reference standard (ICC 0.996, 95% CI: 0.994–0.997).

Conclusion

The AI model attained high diagnostic accuracy for pneumoperitoneum on abdominal CT scans, promising to expedite emergency workflows.

Key Points

Question Reliable AI detection of pneumoperitoneum, particularly for small-volume free air, on emergency CT remains an unmet need for rapid and accurate emergency triage.

Findings The AI model shows high sensitivity and specificity for clinically relevant pneumoperitoneum volumes, although trace-volume detection on CT scans remains challenging.

Critical relevance statement This study evaluates the diagnostic performance and volume-dependent variability of an AI model for pneumoperitoneum detection on CT scans, with the potential to aid emergency radiology workflow prioritization and decision support, though prospective studies are needed to confirm clinical impact.

Graphical Abstract