Automatic generation of pedicle contours in 3D vertebral models

Abstract

Objectives: Pedicle location and recognition play important roles in spinal morphology analysis and orthodontic screw implantation, which can help doctors avoid injuring the pedicle during screw implantation. However, because of the complex spatial structures of vertebrae and the close connection between the pedicle and other parts of the vertebrae, it is challenging to locate and recognize the pedicle of the vertebral arch in 2D or 3D vertebral images.

Methods: In this paper, based on deep learning technology, we propose a method for automatically recognizing the vertebral pedicle in individual vertebral models and drawing pedicle contours. The goal is to provide references so doctors can simulate the pedicle screw implantation operation to prevent screw deviation and further enhance the automation of our team's scoliosis-correction assistive system. First, we preprocess the individual vertebral models to obtain their point clouds. Then, we use a modified PointNet model to segment the pedicle areas from the individual vertebral point clouds. We use the segmentation results to automatically fit the cross-sections of pedicles and finally generate the pedicle contours as surgical references.

Results: The experiments show that the method can generate contours quickly and accurately with a small amount of manual adjustment and can provide good references for simulating screw placement.

Conclusions: The efficiency of generating pedicle contours during the process of simulated screw placement is greatly improved, and the difficulty of using our simulation system has also been greatly reduced, both of which play essential roles in pedicle screw implantation and the formulation of surgical plans.

Keywords: Deep learning; Pedicle contours; Pedicle recognition; Point clouds; PointNet.