Segmentation of Aorta with Aortic Dissection based on Centerline and Boundary Distance

Segmentation of the aorta is important for the diagnosis and treatment of aortic disease. However, low image contrast and blurred boundaries between the aortic region and surrounding tissues can significantly affect segmentation performance. Based on 3D-UNet with spatial attention module, this paper proposes a multi-branch shape-aware segmentation network named CDM-Net, which transforms the traditional segmentation problem into a regression problem of distance transformation map and centerline heatmap. A new inference method based on regression is also proposed, the prediction of our network can be combined with the predictions of other networks. Without changing other segmentation metrics (Dice, ASD), the clDice of the combined method improves by 1.5%. Our proposed method can improve the connectivity of aorta segmentation results, paving the way for accurate centerline extraction and multiplanar reconstruction in the future.

Automated summary

In this paper, a multi-branch shape-aware segmentation network named CDM-Net is proposed based on 3D-UNet with spatial attention module. The traditional segmentation problem is transformed into a regression problem of distance transformation map and centerline heatmap. A new inference method based on regression is also introduced. Without changing other segmentation metrics, the proposed method improves the connectivity of aorta segmentation results.