Automated labeling using tracked ultrasound imaging: Application in tracking vertebrae during spine surgery

Area of Science:

• Medical Imaging
• Machine Learning
• Surgical Navigation

Background:

• Machine learning (ML) requires large labeled datasets for ultrasound (US) image analysis, which are time-consuming to create and prone to labeling errors.
• Automated labeling of US images is crucial for developing robust ML solutions in interventional procedures.

Purpose of the Study:

• To present a practical method for automated labeling of US images by transferring labels from 3D diagnostic images (CT/MR) using tracked US imaging.
• To support supervised training of ML models for tasks like spinal vertebrae segmentation.

Main Methods:

• Tracked US imaging was used to map 3D structures from CT volumes onto 2D US frames.
• A U-Net model (TernausNet) was trained using different labeling strategies (full vertebra, posterior surface, weighted posterior surface).
• Automated data cleaning and augmentation were applied to enhance robustness.

Main Results:

• The weighted posterior surface labeling strategy improved segmentation accuracy (3.86 mm RMSD) compared to direct CT label mapping.
• Registration accuracy was high, with 1.52 mm TRE for the weighted posterior surface strategy.
• Automated data cleaning and augmentation significantly improved segmentation and registration accuracy.

Conclusions:

• An automated labeling method for US imaging was developed, enabling ML model training through 3D-to-2D label transfer.
• Proper probe calibration, data cleaning, and specific labeling strategies are critical for mitigating errors.
• Real-time US imaging shows potential for precise anatomical tracking in surgery.