Automatic segmentation of bone surfaces from ultrasound using a filter-layer-guided CNN

Area of Science:

• Medical Imaging
• Orthopedic Surgery
• Artificial Intelligence

Background:

• Ultrasound (US) is a safe, real-time imaging modality for computer-assisted orthopedic surgery (CAOS).
• Interpreting bone US data is challenging due to noise, artifacts, and low-resolution bone boundaries.
• Robust and efficient segmentation is essential for US-guided CAOS.

Purpose of the Study:

• Develop a convolutional neural network (CNN)-based technique for segmenting bone surfaces from in vivo US scans.
• Improve segmentation robustness against imaging artifacts and low-intensity boundaries.
• Create a computationally inexpensive segmentation mechanism for US-guided CAOS.

Main Methods:

• Utilized a CNN architecture with feature map fusion.
• Employed multi-modal inputs: B-mode US images and local phase filtered images.
• Investigated various fusion architectures for combining image modalities.

Main Results:

• Achieved an average F-score above 95% on 546 in vivo scans.
• Demonstrated an average bone surface localization error of 0.2 mm.
• Results showed statistically significant improvement over state-of-the-art methods.

Conclusions:

• The proposed segmentation method shows promise for CAOS applications due to its accuracy and robustness.
• Further validation is needed to establish the full clinical utility of the method.