Artificial intelligence-enhanced multimodality imaging of the aortic root and ascending aorta
G Imperatore, R E S Shaw, L S Sun, H J JabagiAbstract
Background
Transthoracic echocardiography (TTE) systematically underestimates aortic dimensions compared with computed tomography (CT), potentially influencing surveillance intervals, surgical referral thresholds, and overall risk stratification in patients with aortic root (AoR) and ascending aorta (AAo) dilatation. Given the increasing reliance on multimodality imaging in contemporary aortic disease management, systematic inter-modality discrepancies may introduce clinically relevant misclassification. Artificial intelligence (AI)–based harmonization offers a strategy to reduce measurement bias, align TTE-derived diameters with CT reference standards, and improve identification of patients at risk of clinically significant discordance.
Methods
We retrospectively analyzed 512 patients with paired transthoracic echocardiography and computed tomography measurements of the AoR and AAo. Patient-level splitting was used to derive independent training, validation, and test cohorts. Machine learning models were developed to predict CT-equivalent diameters and to identify clinically relevant inter-modality discordance (>2 mm).Agreement improvement was quantified by mean absolute error (MAE), root mean square error (RMSE), and proportion of measurements within ±2 mm. Statistical comparisons between unadjusted and AI-harmonized TTE were performed using paired bootstrap resampling and McNemar’s test. Robustness and generalizability were evaluated through 5-fold patient-level cross-validation and calibration analysis.
Results
On the independent test set, XGBoost-based harmonization significantly improved agreement compared with unadjusted echocardiography. For AoR, mean absolute error decreased from 2.89 mm to 0.64 mm and the proportion of measurements within ±2 mm increased from 57.6% to 94.6% (p<0.001). AI harmonization reduced guideline-threshold misclassification, improving concordance across clinically relevant diameter cut-offs. For AAo, MAE was reduced from 3.15 mm to 0.94 mm, with within-±2 mm agreement increasing from 68.1% to 81.9% (p<0.001). Improvements in MAE and RMSE were statistically significant by paired bootstrap analysis, and cross-validation demonstrated stable performance across patient subsets. The discordance prediction model showed excellent discrimination (AUC 0.841 for AoR and 0.910 for AAo) with good calibration and stable performance across cross-validation folds.
Conclusions
AI-driven harmonization substantially reduces systematic bias between TTE and CT aortic measurements and enables reliable identification of high-risk discordant cases. This framework supports more accurate multimodality integration, improves measurement consistency across imaging platforms, and may contribute to safer and more precise clinical decision-making in patients with thoracic aortic disease. This approach may facilitate guideline-consistent threshold assessment and improve safety in longitudinal aortic surveillance.AI harmonization of echo aortic diameters.