Joined fragment segmentation for fractured bones using GPU-accelerated shape-preserving erosion and dilation

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Biomedical Engineering

Background:

• Manual segmentation of fractured bones from CT scans is time-consuming and requires extensive radiologist interaction.
• Existing automated methods often struggle with joined bone fragments, necessitating further manual correction.

Purpose of the Study:

• To develop a GPU-accelerated 3D segmentation framework to automate the segmentation of joined bone fragments from CT images.
• To reduce the time cost and interaction required for fractured bone segmentation.

Main Methods:

• A novel framework utilizing normal-based erosion to separate joined fragments, followed by connected component labeling (CCL) for fragment identification.
• Record-based dilation is employed to restore original bone shapes, with a random walk algorithm addressing challenging cases.
• The entire process is accelerated using parallel processing on a graphics processing unit (GPU).

Main Results:

• The framework achieves highly accurate fragment segmentations with Dice scores exceeding 99%.
• Segmentation of a typical fractured bone volume (512 × 512 × 425 voxels) averages only 3.47 seconds.
• The method effectively segments joined bone fragments with minimal user interaction for most cases.

Conclusions:

• The proposed GPU-accelerated framework significantly enhances the efficiency and accuracy of fractured bone segmentation from CT images.
• This automated approach alleviates the burden on radiologists, offering a faster and more precise alternative to manual segmentation.
• The integration of normal-based erosion, CCL, record-based dilation, and a random walk algorithm provides a robust solution for complex fracture segmentation.