Automated measurement of ascending aortic diameters and detection of thoracic aortic aneurysm with a deep learning approach on routine chest radiographs

Abstract Background Because thoracic aortic aneurysm (TAA) poses a significant risk of dissection or rupture, making timely detection essential for preventing complications. While computed tomography (CT) and magnetic resonance imaging provide accurate measurements, their high cost, limited availability, and radiation exposure hinder widespread screening. Chest radiography (CXR) is more accessible but lacks sensitivity in detecting TAA. We developed a deep learning (DL) algorithm to estimate the aortic diameter and detect ascending TAA on CXR. Methods In this retrospective study, we trained a conditional generative adversarial network (Pix2PixHD) on 63,695 CT scans to generate digitally reconstructed radiographs, from which an ascending aorta segmentation model was developed. Post-processing included skeletonization for centerline detection and diameter measurement. Internal validation used 200 scans. For external validation (n=16,488), AI-derived diameters on routine CXRs were compared with manual CT measurements in both axial and coronal views. Results Internal validation showed a structural similarity index measure of 0.9775±0.0020 for aortic segmentation and a mean absolute percentage error (MAPE) of 6.44% for diameter measurement. In external validation, AI-derived measurements correlated strongly with CT-derived diameters, yielding correlation coefficients of 0.787 (95% CI: 0.779–0.796) for coronal CT and 0.737 (95% CI: 0.728–0.745) for axial CT. The mean absolute error was 0.436 cm for coronal and 0.355 cm for axial measurements. The model achieved an area under the receiver operating characteristic curve of 0.842–0.939 at clinically relevant thresholds (≥4.0 cm, ≥4.5 cm, and ≥5.0 cm), demonstrating high sensitivity and specificity for TAA detection. At a prevalence of 0.16%, the negative predictive value exceeded 99.9%, indicating strong rule-out capability. Conclusions Our DL-based approach enables automated, accurate measurement of ascending aortic diameters on routine CXRs, facilitating TAA screening. If validated further across diverse populations and healthcare settings, this cost-effective DL tool could augment current diagnostic pathways, reduce reliance on advanced imaging, and potentially improve early detection and management of TAA.