Deep-learning approach to automate the segmentation of aorta in non-contrast CTs.

Purpose: Segmentation of vascular structures in preoperative computed tomography (CT) is a preliminary step for computer-assisted endovascular navigation. It is a challenging issue when contrast medium enhancement is reduced or impossible, as in the case of endovascular abdominal aneurysm repair for patients with severe renal impairment. In non-contrast-enhanced CTs, the segmentation tasks are currently hampered by the problems of low contrast, similar topological form, and size imbalance. To tackle these problems, we propose a novel fully automatic approach based on convolutional neural network. Approach: The proposed method is implemented by fusing the features from different dimensions by three kinds of mechanisms, i.e., channel concatenation, dense connection, and spatial interpolation. The fusion mechanisms are regarded as the enhancement of features in non-contrast CTs where the boundary of aorta is ambiguous. Results: All of the networks are validated by three-fold cross-validation on our dataset of non-contrast CTs, which contains 5749 slices in total from 30 individual patients. Our methods achieve a Dice score of 88.7% as the overall performance, which is better than the results reported in the related works. Conclusions: The analysis indicates that our methods yield a competitive performance by overcoming the above-mentioned problems in most general cases. Further, experiments on our non-contrast CTs demonstrate the superiority of the proposed methods, especially in low-contrast, similar-shaped, and extreme-sized cases.