Whole Spine Segmentation Using Object Detection and Semantic Segmentation.

Area of Science:

• Medical Imaging
• Computer Vision
• Spine Surgery

Background:

• Virtual and augmented reality are increasingly used in spine surgery for planning, training, and navigation.
• Accurate identification of osseous structures is crucial for 3D virtual reconstructions.
• Manual labeling of vertebrae on CT scans is time-consuming.

Purpose of the Study:

• To develop a fully automated pipeline for vertebra segmentation on computed tomography (CT) scans.
• To provide a foundation for virtual/augmented reality applications and radiomic analysis in spine surgery.
• To automate the process of labeling vertebrae, reducing manual effort.

Main Methods:

• A two-stage approach was employed, starting with vertebra detection using YOLOv8m.
• Semantic segmentation was then performed on cropped images using a 2D-U-Net architecture.
• Models were trained on a large public dataset and externally validated on two additional datasets.

Main Results:

• The object detection model achieved a mean average precision (mAP50) over 0.84 for vertebra recognition.
• The segmentation model demonstrated robust performance with a mean Dice score of 0.75 (internal) and up to 0.82 (external validation).
• The pipeline was trained and validated on 214 and 40 CT scans, respectively.

Conclusions:

• A robust, automated two-stage pipeline for vertebra detection and segmentation in CT scans was successfully developed.
• The proposed method shows strong performance in identifying and delineating bony structures, suitable for clinical applications.
• This automated approach can significantly streamline workflows in virtual reality-assisted spine surgery and radiomic studies.