Multi-perspective region-based CNNs for vertebrae labeling in intraoperative long-length images

Area of Science:

• Medical Imaging
• Artificial Intelligence in Surgery
• Spine Surgery Imaging

Background:

• Intraoperative imaging is crucial for surgical guidance.
• Automated analysis of multi-angle x-ray images can enhance surgical procedures.
• Accurate vertebrae labeling is essential for spine surgery to prevent errors.

Purpose of the Study:

• To develop and evaluate multi-perspective region-based neural networks for automatic vertebrae labeling in Long-Film images.
• To leverage imaging geometry for improved automated image analysis during surgery.

Main Methods:

• Designed a multi-perspective network architecture exploiting small view-angle disparities for information consolidation.
• Incorporated a second network for joint labeling using large view-angle disparities (AP and lateral views).
• Utilized a recurrent module for contextual information and anatomical order enforcement, forming the multi-view multi-slot (MVMS) network.

Main Results:

• The MVMS network achieved >96% F1 score and 88.3% labeling accuracy, outperforming conventional methods.
• Aggregation of multiple vertebral appearances significantly improved labeling accuracy (p < 0.05).
• Demonstrated robustness to surgical instrumentation, with comparable performance in images with and without it.

Conclusions:

• The MVMS network enables accurate, automated vertebrae labeling during spine surgery through effective multi-perspective aggregation.
• The algorithms show potential for generalization to other multi-view imaging tasks and modalities.
• Automated labeling can prevent adverse events, facilitate image registration, and enable automated spinal alignment measurements.