Fragment distance-guided dual-stream learning for automatic pelvic fracture segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Orthopedic Surgery

Background:

• Pelvic fractures are severe injuries requiring precise segmentation of bone fragments from CT scans for diagnosis and treatment planning.
• Current segmentation methods face challenges due to the variability in fracture fragment morphology, quantity, and location.

Purpose of the Study:

• To develop a novel dual-stream learning framework for automatic segmentation and category labeling of pelvic fractures.
• To improve the accuracy and efficiency of preoperative planning for pelvic fracture management.

Main Methods:

• A dual-stream learning framework with a dual-branch architecture leveraging distance learning from bone fragments.
• Implementation of a multi-size feature fusion module for adaptive aggregation of features from diverse receptive fields.
• Extensive experiments conducted on three distinct pelvic fracture datasets from multiple medical centers.

Main Results:

• The proposed method achieved a mean Dice coefficient of 0.935±0.068 and mean Sensitivity of 0.929±0.058 on the FracCLINIC dataset.
• On the FracSegData dataset, the method obtained a mean Dice coefficient of 0.955±0.072 and mean Sensitivity of 0.912±0.125.
• Performance metrics demonstrated superiority over existing comparative methods, highlighting accuracy and generalizability.

Conclusions:

• The developed dual-stream learning framework effectively addresses the challenges of pelvic fracture segmentation.
• The method optimizes pelvic fracture segmentation, offering a potentially valuable tool for preoperative planning in clinical settings.