Deep learning-based detection of ascending aortic dilatation on chest radiographs: A diagnostic study

This study externally tests the performance of an artificial intelligence algorithm (AI) for diagnosing ascending aortic dilatation (AAD) using PA view chest radiography (PA CXR).

Two retrospectively collected cohorts with paired CXR/CT within 30 days (Group 1) and 90 days (Group 2) were gathered as external test sets. The performance of AI (DeepCatch X Aorta v1.1.0) for detecting AAD using PA CXR was analyzed against semi-automatic measurements on CT using fixed thresholds (4.0, 4.5, and 5.0 cm) and subjective assessment (dilatation or no dilatation) of 2 cardiovascular radiologists (>20 years' experience) and 3 less-experienced (<2 years' experience) doctors. Additionally, modified thresholds accounting for radiographic magnification were analyzed in the Group 1. Performance metrics (sensitivity, specificity, and AUC) and inter-reader agreement were used.

Group 1 included 336 scans (mean age = 64 years) and Group 2 had 190 scans (mean age = 70 years). Dilatation (≥4 cm) was present in 182 patients (Group 1) and 94 patients (Group 2). The algorithm demonstrated robust performance in both Groups for the ≥4 cm threshold, with AUC of 0.89 (95 % CI: 0.85, 0.92) in Group 1 and 0.88 (95 % CI: 0.83, 0.92) in Group 2. All readers' performances were under the AUC of AI at 4 cm threshold. Inter-reader agreement was fair to moderate (Cohen's kappa, Group 1, 0.38; Fleiss's kappa, Group 2, 0.37).

The AI algorithm detected AAD from a single PA CXR with robust AUC over 0.8 across different clinical thresholds and cohorts, outperforming human readers.