- Keywords: CAC, Calcium Score, Cardiac CT, Deep Learning, Semantic Segmentation
- TL;DR: Calcium Score prediction and CAC localization using anisotropic convolutional networks.
- Abstract: Abstract—In this work we propose to apply deep learning semantic segmentation techniques to calcium quantification and localization. 3D CT chest imaging is an essential support to diagnosis of cardiovascular disease and coronaires calcification burden is one of its strongest indicators. CAC is quantified using a per coronary branch Agatston score. In this clinical context using deep learning techniques for multi class segmentation we could design an algorithm that automatically localizes and quantifies calcifications volumetry and Agatston score. The architecture used is inspired by Vnet , a popular model adapted to this particular CT exam modality, the key contribution is the use of anisotropic pooling and unpooling layers. 124 patients were provided by ***** and manually annotated by experts with clinical feedback. As a result we could achieve 0.9 average R2 = 1 - rMSE (relative mean square error) on multiple branches on a test set of 14 patient left out from the whole dataset. Index Terms—CAC, Calcium Score, Cardiac CT, Deep Learning, Semantic Segmentation.
- Track: short paper
- Paper Type: well-validated application