A Multi-task Network with Centerline Supervision for 3D Pelvis Artery Segmentation on CECT images

Accurate pelvis artery segmentation on computerized tomography images is crucial for many diagnoses and surgery related to abdominal diseases, like colorectal cancer metastasis analysis. However, seldom researches considered this, and many minor branch vessels are always broken or even disappear in existing segmentation methods, since they are thin and long, with low contrast, and many variants may exist. To address this problem, we provide a multi-task network with centerline supervision. The centerline extracting task is introduced to offer other domain-specific information and help the network learn the branch vessels. This method mainly includes an early-branch network with an inter-task feature fusion module. In the proposed network, the hard-sharing encoder extracts the shared spatial and global information. The dual-path decoder separates the learning of the tasks and has some fusion blocks to combine and exchange inter-task features. Then, a multi-task learning loss with a consistency correction penalty is designed to keep the training balance through punishing the gap between the two tasks. To evaluate the performance of our method, we conduct ablation and comprehensive experiments on a local pelvis artery dataset. Experimental results in metrics and visualization show that the proposed method achieves superior performance.