Aortic Stenosis (AS) is a valvular heart disease characterized by obstructed blood flow from the left ventricle to the aorta. This condition is primarily caused by structural abnormalities of the aortic valve, particularly calcification, which prevent...
Aortic Stenosis (AS) is a valvular heart disease characterized by obstructed blood flow from the left ventricle to the aorta. This condition is primarily caused by structural abnormalities of the aortic valve, particularly calcification, which prevents the leaflets from opening properly. Its prevalence increases sharply with age, and the prognosis deteriorates rapidly upon symptom onset, posing a significant health challenge in an aging society.
The current diagnostic standard, Transthoracic Echocardiography (TTE), while non-invasive and widely used, suffers from significant operator-dependency, leading to inter-observer variability in results. It also has inherent limitations, such as reduced image quality in obese patients or those with pulmonary disease. As an alternative, cardiac CT can measure the Aortic Valve Calcium (AVC) score; however, this score merely quantifies the total calcium burden and fails to capture crucial morphological characteristics linked to severity, such as the 3D distribution, density, and geometry of the calcification.
To overcome these limitations and achieve objective, accurate automated classification of AS severity, this study proposes a Artificial Intelligence model that fuses quantitative radiomics features with high-dimensional deep learning features extracted from cardiac CT images. The proposed model automatically segments the AVC region and, based on this segmented region, crops a 3D Region of Interest (RoI) critical for classification. From this RoI, quantitative radiomics features and high-dimensional deep learning features (based on a 3D deep learning encoder) are extracted in parallel. These two heterogeneous feature vectors are then effectively fused using a Gated Fusion mechanism. To prevent overfitting, the fused features undergo Lasso-based feature selection before being fed into a Logistic Regression classifier for final classification.
Experiments conducted on data from 406 patients demonstrated that the proposed model, which utilizes AVC RoI cropping, Gated Fusion, Lasso, and Logistic Regression, achieved the highest performance with an Accuracy of 0.8148, F1-Score of 0.7706, AUROC of 0.9226, and AP of 0.8007. This result validates that the quantitative information from radiomics and the abstract spatial information from deep learning are complementary, and that the Gated Fusion mechanism effectively combines these heterogeneous features. The automated classification model proposed in this study is expected to assist clinicians in making objective diagnoses, thereby reducing diagnostic time and alleviating the burden of interpretation.