Integrating 3D Geometry of Organ for Improving Medical Image Segmentation

Prior knowledge of organ shape and location plays an important role in medical imaging segmentation. However, traditional 2D/3D segmentation methods usually operate as pixel-wise/voxel-wise classifiers where their training objectives are not able to incorporate the 3D shape knowledge explicitly. In this paper, we proposed an efficient deep shape-aware network to learn 3D geometry of the organ. More specifically, the network uses a 3D mesh representation in a graph-based CNN which can handle the shape inference and accuracy propagation effectively. After integrating the shape-aware module into the backbone FCNs and jointly training the full model in the multi-task framework, the discriminative capability of intermediate feature representations is increased for both geometry and segmentation regularizations on disentangling subtly correlated tasks. Experimental results show that the proposed network can not only output accurate segmentation, but also generate smooth 3D mesh simultaneously which can be used for further 3D shape analysis.