RASNet: U-Net-Based Robust Aortic Segmentation Network for Multicenter Datasets

Abstract

The segmentation and reconstruction of the aortic vessel tree (AVT) is necessary in detecting aortic diseases. Currently, the mainstream method must be deployed manually, which is time-consuming and requires an experienced radiologist/physician. Automatic segmentation methods developed in recent years have performed well on single-centered datasets. However, their performance degraded on multi-centered datasets due to the various specifications of the data. We propose a 3D U-Net-based robust aortic segmentation framework to address the problem. We implied Hounsfield Units (HU) adaptive method during preprocessing to reduce the variety of intensity distribution of the inter-center images. We insert convolutional block attention modules (CBAM) in our network to improve its channel and spatial representation ability. Furthermore, we set a two-stage training process and introduce the Hausdorff distance (HD) loss in the second stage to optimize the structure of the segmentation results. Using a specific validation set collected from the multicenter AVT dataset which includes samples D5, D6, K4, K5, R5, R6, our proposed method reached an average Dice Similarity Coefficient (DSC) of 0.9396 and an average HD of 16.1.