Vision transformer-based multi-class classification of aortic calcification scores

Abstract

This study presents a fully ultrasound-based and cost-effective approach for the automatic grading of aortic valve calcification, which plays a critical role in the assessment of aortic stenosis. To eliminate radiation exposure associated with computed tomography, the proposed method relies exclusively on echocardiographic images and was trained on a dedicated dataset constructed for this study. A Vision Transformer (ViT) model operating on ROI frames extracted from the aortic valve region was developed to classify four calcification levels: mild, moderate, severe, and critical. During testing, the model achieved 85% accuracy and a macro-averaged precision of 0.74, demonstrating a reliable decision mechanism with a low false-positive tendency. To further evaluate the architectural choice, ImageNet pre¬trained ResNet50 and EfficientNet-B3 models were trained and tested under identical conditions. Although CNN-based architectures produced competitive results, the ViT model demonstrated more balanced performance, particularly in intermediate and advanced grades, and achieved the highest macro-averaged metrics among the evaluated models. By providing a radiation-free, reproducible, and economically accessible solution, the proposed framework offers a clinically applicable decision-support system for the evaluation of aortic stenosis.