Assessment of Thoracic Aortic Morphology from 3D CT Scan Images Using a Multi-Stage Machine Learning Model for Representation Learning, Clustering, and Disease Prediction.

Abstract

Abstract Introduction Aortic diseases (AD) are associated with significant morphological changes along the thoracic aorta (TA). Thoracic aortic aneurysm and dissection are among the most common morphological manifestations of AD, representing two major life-threatening conditions that contribute eventually to morbidity and mortality if left undetected and untreated. Morphological changes can arise from either inherited connective tissue disorders, such as Marfan syndrome (MFS) or natural aging that can cause different morphological patterns which not fully identified. Current risk assessment of the AD relies in 1 dimensional measurement of the aortic size which has been shown as a poor risk predictor. Therefore, there is a need for more accurate and advanced assessment tool to identify the variations of the morphological changes in MFS and aging groups. Recently, a growing subfield of machine learning (ML), particularly self-supervised machine learning (SSML), has gained attention for its ability to extract clinically meaningful anatomical patterns from 3-dimensional (3D) medical imaging without requiring large, labelled datasets. Despite numerous studies on ML in medical imaging, the application of SSML in these patients remains due to the insufficient availability of large, population-based imaging datasets. Aims (1) to generate an accurate images of volume-rendered thoracic aortas from patients with MFS, age-related changes; (2) to develop a SSML model that capable of extracting and learning morphological patterns in MFS and ageing; (3) to identify meaningful clusters of different aortic morphologies; and (4) to evaluate the added value of SSML pretraining by assessing the improvement in computer vision models for predicting MFS status from aortic volume-rendered images. Research question 1. Can a self-supervised machine learning model pre-trained on 3D thoracic aorta images capture the different morphological patterns Marfan syndrome and aging? 2. Can a predictive model distinguish between Marfan and non-Marfan cases across the entire cohort? Method A total of 117 3D volume-rendered images from patients with MFS (pre- and post-operative), aging group, and control group were used for model training and evaluation. Hierarchical agglomerative clustering (HAC) was applied to SSL-derived embeddings to explore latent morphological subgroups, while a predictive model was developed to classify Marfan vs. non-Marfan cases, comparing performance with and without SSML pretraining. For categorical variables (e.g., gender, ethnicity), one-vs-all chi-squared tests were used. For continuous variables (e.g., age, aortic measurements), one-vs-all Welch’s t-tests were applied, and results were summarized as mean ± standard deviation (SD). Results The SSML framework successfully captured discriminative morphological pattens of the TA represented as embeddings of the Uniform Manifold Approximation and Projection (UMAP). The HAC revealed 6 subclusters that were clinically meaningful clusters reflecting variations across all the measured anatomical parameter in sinus of Valsalva diameters, ascending aortic length, and overall thoracic morphology. Control-like clusters were consistently grouped in clusters 6.1 and 6.2. MFS patients grouped into clusters, 6.3, 6.5, and 6.6. whereas cluster 6.4 represented individuals with age-related changes. Predictive evaluation demonstrated that the SSML-pretrained model outperformed the baseline, achieving higher accuracy (90.6%), sensitivity (87.9%), specificity (93.2%), and area under the ROC curve (AUC 0.97 vs. 0.84).

Conclusion The SSML framework developed in this study demonstrated strong performance in detecting aortic pathologies and shows promise in supporting physicians with the early identification of patients with Marfan syndrome. By enabling more accurate recognition of characteristic morphological patterns, this approach could ultimately support diagnostic decisions and improve patient outcomes. These methods represent excellent candidates for advancing state-of-the-art prediction model of Marfan syndrome based solely on imaging features.