Deep Learning-Based Segmentation of Gravity-Loaded Human Spine.

Area of Science:

• Biomedical Engineering
• Radiology
• Artificial Intelligence

Background:

• Accurate spinal alignment analysis is crucial for musculoskeletal disorders.
• Traditional imaging lacks 3D load-bearing spinal data.
• Weight-bearing CBCT offers improved spinal imaging.

Purpose of the Study:

• To develop a deep learning protocol for segmenting weight-bearing CBCT images.
• To enable precise 3D analysis of spinal alignment under gravity.
• To facilitate accurate diagnosis and management of spinal conditions.

Main Methods:

• Utilized a U-Net convolutional neural network (CNN) with 3D convolutional layers and residual connections.
• Developed a protocol encompassing image acquisition, manual annotation, preprocessing, and model training.
• Focused on segmenting vertebral bodies, pelvis, and femoral head from weight-bearing CBCT scans.

Main Results:

• Successfully segmented key anatomical structures in gravity-loaded spinal images.
• Enabled accurate measurement of clinical parameters like Cobb angle and vertebral rotation.
• Generated 3D models suitable for 3D printing for surgical planning and education.

Conclusions:

• The deep learning protocol provides a reliable method for segmenting spinal structures in weight-bearing conditions.
• This approach enhances diagnostic accuracy for spinal disorders.
• The protocol is adaptable for segmenting other anatomical structures under load, with broad clinical and research applications.