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

Area of Science:

• Medical Imaging
• Computer-Aided Surgery
• Biomedical Engineering

Background:

• Accurate 3D segmentation of lower-limb bones (tibia, fibula) from computed tomography (CT) is crucial for surgical planning and navigation.
• Large-scale volumetric CT data presents computational challenges and risks overlooking long-range spatial voxel connections.
• Existing methods may struggle with efficiency and capturing global context in 3D bone segmentation.

Purpose of the Study:

• To develop an accurate and efficient 3D segmentation approach for lower-limb bones in volumetric CT.
• To address computational costs associated with large 3D datasets.
• To improve the capture of long-range spatial relationships between voxels for enhanced segmentation.

Main Methods:

• Proposed an efficient segmentation framework utilizing 2D projection view-based slice filtering for large-scale volumetric CT.
• Implemented parameter-reduced separable convolution to optimize computational efficiency.
• Developed a novel voxel group attention mechanism to emphasize long-range voxel group connections and improve network representation.

Main Results:

• The proposed 3D bone segmentation approach achieved high accuracy with limited computational resources.
• Demonstrated superior performance compared to state-of-the-art 3D models for lower-limb bone segmentation.
• Effectively addressed the challenges of computational cost and long-range spatial dependencies in CT data.

Conclusions:

• The developed method offers an accurate and efficient solution for 3D lower-limb bone segmentation from volumetric CT.
• The approach is suitable for surgical planning and navigation applications requiring precise bone models.
• The voxel group attention mechanism significantly enhances the segmentation network's ability to capture global context.