CSR-Net: Cross-Scale Residual Network for multi-objective scaphoid fracture segmentation

Area of Science:

• Orthopedic surgery
• Medical imaging analysis
• Artificial intelligence in healthcare

Background:

• Scaphoid fractures are common and challenging to heal due to the bone's anatomy.
• Accurate 3D reconstruction of scaphoid fractures is crucial for surgical planning, especially screw placement.

Purpose of the Study:

• To introduce a novel deep learning model, the cross-scale residual network (CSR-Net), for precise scaphoid fracture segmentation.
• To enable synchronous segmentation of both scaphoid fractures and surrounding hand bones.

Main Methods:

• Development of the CSR-Net utilizing cross-scale residual connections for feature fusion across different layers.
• Implementation of a multi-objective architecture in the output layer and channel design.
• Testing the model on 65 computed tomography (CT) images of scaphoid fractures.

Main Results:

• CSR-Net demonstrated superior performance in segmenting both hand bones and scaphoid fractures compared to existing methods.
• Visually, the segmentation provided clearer and more intuitive displays of the fracture surface.
• The model achieved accurate and rapid segmentation of scaphoid fractures.

Conclusions:

• The CSR-Net offers an effective solution for accurate and rapid scaphoid fracture segmentation.
• Its multi-objective design facilitates the creation of precise digital models for surgical navigation in hand bone procedures.