Beyond gas bubbles: AI analysis of the "bubble bed" microenvironment improves diagnosis of infected abdominal collections.

Authors

Abstract

Accurate diagnosis of infected intra-abdominal fluid collections (IAFCs) is challenging, as the conventional "gas bubble sign" on computed tomography (CT) has poor sensitivity. This study aimed to develop and validate a fully automated artificial intelligence (AI) model using non-contrast CT to improve diagnostic accuracy. In this multicenter retrospective study (July 2011-July 2024), 797 patients with IAFCs confirmed by culture were divided into training (n = 637), validation (n = 80), and external test (n = 80) sets. We developed an AI model, Bubble Bed Based Learning Engine for Abdominal Infection (BUBBLE-AI), based on the novel "bubble bed" concept, which analyzes the inflammatory microenvironment around gas bubbles. The model integrates deep learning and radiomic features, extracted from automated segmentations, with clinical data. The BUBBLE-AI model demonstrated robust and generalizable performance, achieving an area under the curve (AUC) of 0.92 in validation and 0.82 (95% CI: 0.72-0.93) in external testing, significantly outperforming traditional methods (P < 0.001). The model achieved a total misdiagnosis rate of 24.1%, a substantial improvement over the bubble sign (38.5%) and a clinical-only model (34.3%). Feature analysis confirmed the "bubble bed" microenvironment was the most dominant source of predictive information (35.8% of features), validating our core hypothesis. The BUBBLE-AI provides a fully automated and accurate tool for diagnosing IAFC infections from non-contrast CT. By identifying infection status more reliably, it can guide antimicrobial stewardship, reduce diagnostic errors, and optimize clinical decision-making.

Tags

Topics