Advancing morphometric assessment of the aorta and left ventricle from dynamic CT: a deep learning-based study

Morphometric analysis of the thoracic aorta (TA) and left ventricle (LV) plays a fundamental role in detecting anatomical abnormalities and functional alterations to support pre-operative planning, predict disease risk and inform device design. However, conventional approaches to morphometric evaluation are typically performed manually using visualization software, thus resulting in time-consuming, operator-dependent processes that are usually limited to static imaging. This work presents an automated three-dimensional image-based methodological framework for dynamic morphometric analysis, from ECG-gated CT datasets.

A multi-label 3D U-Net was trained for the automatic segmentation of the TA and LV using a dataset of 50 single-phase CT scans, with ground-truth label maps validated under expert radiological supervision. Model performance was tested on an independent multi-phase cohort of 10 patients. The network achieved high segmentation accuracy, with Dice scores of 97.77 ± 0.31% for the TA and 91.45 ± 1.26% for the LV on the multi-phase test set. The resulting 3D surface models enabled the computation of geometric descriptors, including volumetric indices, displacement fields, and centreline-based diameters, across cardiac phases on 42 patients. Overall, the framework demonstrated robustness to variations in contrast intensity, cardiac motion, and inter-patient anatomical variability, providing a reliable and reproducible pipeline for comprehensive, three-dimensional, and time-resolved morphometric analysis of the ventriculo-arterial complex with physiological or mildly altered anatomy.

This approach has strong potential for future clinical translation, supporting quantitative assessment of cardiac function, and aortic pathophysiology.