Enhancing accuracy of detecting left atrial dilatation on CT pulmonary angiography

Introduction: Left atrial (LA) dilatation predicts several cardiovascular disorders. Identifying LA dilatation on computed tomography pulmonary angiography (CTPA) could aid diagnosis of cardiovascular disease. This study assessed an artificial intelligence (AI) segmentation model's performance at detecting LA dilatation on CTPA. Methods: Patients with suspected pulmonary hypertension (PH) who underwent CTPA and cardiac MRI (CMR) were retrospectively identified from a single centre registry. The LA was segmented by an AI tool for CTPA and a validated AI tool for CMR. LA volume measurements were categorised for LA dilatation based on existing threshold values. The expert radiologist's reports of the CTPA studies were also categorised for LA dilatation. Automated CTPA LA volumes and corresponding radiologist reports were compared against the reference standard of CMR. Results: 451 patients were included (mean age 64 ± 13 years, 62.5 % female, 85.8 % white). Automated LA volume measurements on CTPA showed strong positive correlation with those on CMR (ρ = 0.92, p < 0.001) with minimal bias on Bland-Altman analysis (-4 mL, 95 %CI -39 to +31 mL). Automated LA measurements on CTPA showed higher agreement with those on CMR (κ = 0.80) than the radiologist reports (κ = 0.62). Automated LA measurements on CTPA showed higher accuracy metrics (sensitivity 81.0 %, specificity 96.8 %, positive predictive value (PPV) 88.5 %, negative predictive value (NPV) 94.4 %) than the radiologist reports (sensitivity 66.7 %, specificity 93.1 %, PPV 74.5 %, NPV 90.2 %). Conclusion: Deep learning increases the accuracy of LA volume measurements on non-ECG gated CTPA, improving radiologist performance in detecting LA dilatation.