Deep Learning-Based Segmentation and Mesh Reconstruction of the Aortic Vessel Tree from CTA Images

Theodoros P. Vagenas, Konstantinos Georgas, George K. Matsopoulos

Abstract: Segmentation of the Aortic Vessel Tree (AVT) in the Computed Tomography Angiography (CTA) images is pivotal for the diagnosis and monitoring of the aortic diseases. Identifying changes in the AVT structure requires high-quality reconstructions that can enable the accurate comparison of the AVT geometry between follow-up scans. However, manual delineation of the whole AVT is a very time-consuming and labor-intensive procedure that can stall the clinical workflow. In this paper, a Convolutional Neural Network (CNN) methodology is implemented based on the SegResNet architecture for the automatic segmentation of the AVT. A training scheme including preprocessing and data augmentation is designed for the memory-efficient and effective learning of the model parameters. Furthermore, reconstructed surfaces from the initially extracted segmentations are produced through the Marching cubes algorithm and surface correction techniques. The proposed methodology is evaluated in the public SEG.A. grand challenge dataset where in a 5-fold cross-validation experiment it achieved DSC coefficient 91.70%, Recall 91.70%, Precision 91.90% and Hausdorff distance 5.17 mm.