A Deep Learning-Based Fully Automated Vertebra Segmentation and Labeling Workflow.

Area of Science:

• Medical Imaging and Artificial Intelligence
• Spine Surgery Navigation
• Deep Learning in Healthcare

Background:

• Spinal disorders are increasingly prevalent, necessitating accurate anatomical analysis for surgical navigation.
• Current manual segmentation methods for spinal structures are inefficient and inconsistent.
• Automated analysis is crucial for high-precision positioning in robotic spine surgery.

Purpose of the Study:

• To develop a fully automated deep learning workflow for vertebral segmentation and labeling.
• To provide efficient and accurate preoperative analysis for spine surgery navigation robots.
• To overcome limitations of traditional manual segmentation methods.

Main Methods:

• Utilized YOLOv7 for 2D vertebral localization on CT sagittal slices, converting the problem to 2D.
• Employed DBSCAN clustering to aggregate 2D detections into 3D vertebral centers, reducing inference time.
• Implemented a 3D U-Net with attention for precise segmentation and a ResNet-Transformer network for labeling.

Main Results:

• Achieved excellent performance in localization (Mean Localization Error: 1.42 mm), segmentation (DSC: 0.968), and labeling (classification accuracy: 94.36%).
• Segmentation metrics included IoU of 0.879, PA of 0.988, MSD of 1.09 mm, and HD of 5.42 mm.
• Validated on 106 spinal CT datasets across diverse clinical scenarios.

Conclusions:

• The developed automated workflow demonstrates high accuracy and efficiency in vertebral localization, segmentation, and labeling.
• This method shows significant potential for enhancing preoperative analysis and navigation support in robotic spine surgery.
• The findings confirm the effectiveness for clinical deployment in spinal surgery navigation systems.