Graph Reasoning and Shape Constraints for Cardiac Segmentation in Congenital Heart Defect

Abstract

Congenital heart defects (CHD) are the most common type of birth defect. The structures of the heart or great vessels of CHD patients have significant variations, which limit the application scope of existing whole heart segmentation methods. To solve this limitation, we propose a CHD segmentation method based on graph reasoning and shape constraints in the study. Graph reasoning is used to capture global relations, through projecting a set of features into the interaction space and then makes relational reasoning. After reasoning, relation-aware features are mapped back to the original feature map. The shape constraints are used to calculate the shape information of each substructure of the heart, this prior knowledge on shape constraints can improve the prediction accuracy. Our method is validated with 68 3D CT datasets with congenital heart defect. The evaluation results show that our method can increase the mean Dice score by 2.1% compared with the state-of-the-art methods. Code: https://github.com/liut969/CHD-Seg.