Efficient lower-limb segmentation for large-scale volumetric CT by using projection view and voxel group attention.

Abstract

3D segmentation of lower-limb bones such as tibia and fibula from volumetric computed tomography (CT) is very important for surgical planning and navigation. However, the input data of the large-scale 3D volumetric CT will increase the computation cost, and meantime, the spatial connections of the long-range voxels will be easily ignored. To solve these problems, this paper proposes an accurate and efficient 3D bone segmentation approach based on the distribution characteristics of the lower-limb bones in volumetric CT. On one hand, the efficient segmentation framework was built for the large-scale volumetric CT using 2D projection view-based slice filtering and parameter-reduced separable convolution. On the other hand, we developed a new voxel group attention mechanism to emphasize the connection of the long-range voxel groups and improve the representational capability of the segmentation network. The experimental results showed that the proposed 3D bone segmentation approach achieved high segmentation accuracy under the conditions of limited computations. It was additionally shown that the proposed approach outperformed the state-of-the-art 3D models for the segmentation of lower-limb bones.