CenterlineNet: Automatic Coronary Artery Centerline Extraction for Computed Tomographic Angiographic Images Using Convolutional Neural Network Architectures

The prevention of cardiovascular diseases starts by a thorough examination of the coronary artery vessels for atherosclerotic plaques existence. By combining deep learning convolutional Neural Network (CNN) architectures and the biological knowledge, we introduce a novel method for the automatic extraction of coronary artery centerlines in Computed Tomography Angiography (CTA) data. The proposed method is based on a 3D convolutional neural network used as a local vessel centerline detector to extract the main and side branches of the coronary artery tree. Coupled with a preprocessing vessel enhancing step, the model takes advantage of the analytical analysis of the former as well as the feature extraction ability of the CNN. In addition, automatic aorta and heart region detections are also performed to delineate the origin and the extent of the coronary tree and phase out the false predictions. We investigate the use of Coronary Stenosis' Detection and Quantification (2012) Database to enrich the training process while ensuring the homogeneity of the scans. Our work offers a fully automatic, CNN based, pipeline for coronary artery centerline detection while circumventing the scarcity of the medical data in hand. Its validation against the ground truth annotation from the Rotterdam Coronary Artery Algorithm Evaluation Framework yields a retrieval ratio of 95% and 93%.