Robust real-time bone surfaces segmentation from ultrasound using a local phase tensor-guided CNN.

Area of Science:

• Medical Imaging
• Computer-Assisted Surgery
• Machine Learning

Background:

• Automatic bone surface segmentation is crucial for computer-assisted orthopedic surgery using ultrasound (US).
• Existing methods struggle with US image variations and dataset bias, limiting their real-world application.
• Manual parameter selection is often required, hindering fully automated workflows.

Purpose of the Study:

• To develop an intensity-invariant convolutional neural network (CNN) for robust bone surface segmentation in ultrasound (US) data.
• To address variations in US imaging artifacts, transducer operation, and machine settings.
• To overcome dataset bias issues prevalent in deep learning-based segmentation methods.

Main Methods:

• Proposed an intensity-invariant CNN architecture for US bone surface segmentation.
• The CNN generates local phase tensor (LPT) and global context tensor (GCT) outputs, invariant to intensity variations.
• LPT and GCT are fused for final segmentation; LPT branch is supervised without manual annotation.

Main Results:

• Evaluated on 1227 in vivo US scans from two different US machines and 28 volunteers.
• Achieved statistically significant improvements in cross-machine bone surface segmentation compared to state-of-the-art methods.
• Demonstrated a computation time of 30 milliseconds per image, a significant speed improvement.

Conclusions:

• The proposed intensity-invariant CNN shows promise for robust bone surface segmentation in US imaging.
• Further validation on diverse US data and exploration of its use in registration methods are planned.
• The method offers a significant advancement for ultrasound-guided computer-assisted orthopedic surgery.