A CNN-based method to reconstruct 3-D spine surfaces from US images in vivo

Area of Science:

• Medical Imaging
• Biomedical Engineering
• Machine Learning

Background:

• Three-dimensional (3-D) spine surface reconstruction is crucial for diagnosing spine disorders and guiding surgery.
• Non-invasive ultrasound (US) imaging offers a promising modality for spine surface visualization.

Purpose of the Study:

• To develop and evaluate a new technique for 3-D lumbar spine surface reconstruction using free-hand ultrasound images.
• To improve the accuracy and density of detected spine surface points compared to existing methods.

Main Methods:

• A late fusion (LF)-based U-net convolutional neural network (CNN) was trained on B-mode and shadow-enhanced B-mode ultrasound images.
• The trained U-net predicted spine surface labels, which were then used for 3-D reconstruction.
• Transducer pose estimation involved registering US image stacks to a CT-derived geometrical model.

Main Results:

• The LF-based U-net increased averaged US surface points by 21.61% and reduced mean absolute error (MAE) by 26.28% compared to a phase symmetry method.
• The overall MAE for spine surface point detection was 0.24±0.29 mm.
• The U-net demonstrated effective detection of the spine posterior arch with high accuracy and point density.

Conclusions:

• The proposed U-net effectively reconstructs 3-D lumbar spine surfaces from ultrasound, offering low MAE and high point density.
• This reconstruction framework can complement or potentially replace external tracking systems in intra-operative spine applications.