Quantification of Abdominal Aorta Calcium Using Convolutional Neural Networks

Arterial calcification is an independent predictor of cardiovascular disease (CVD) events whereas abdominal aorta calcium (AAC) detection using non-enhanced computed tomography (CT) images might help to anticipate extra coronary outcomes. On this work, three single-input convolutional neural networks (CNNs) were trained to classify aortic calcifications candidates using axial, sagittal and coronal patches and therefore automate the evaluation of the AAC score in patients with essential hypertension. Cardiac CT images from 456 patients were analyzed together with their aortic geometry, that was previously assessed. Orthogonal patches centered in each lesion candidate inside and near the abdominal aorta were reconstructed and a dataset with 12,570 images (76.6% positives) was built. The CNN of axial patches had the best performance, reaching a 0.951 F1-score, 95.9% sensitivity and 93.4% of the subjects correctly classified in their AAC category. This work shows that AAC can be successfully classified and quantified using CNN architectures to properly classify patients into different CVD risk categories.