Corporation de l'École Polytechnique de Montréal

Published purpose

The aim of this project is to develop vertebra automatic segmentation and identification tools, in two different modalities: (1) pre-operative MRI scans and (2) per-operative surface scans. For (1), we will exploit existing public (open-source) spine MRI/CT datasets to train a convolutional neural network (CNN) for the task of segmentation. In contrast with existing methods, we will focus on combining local vertebra receptive fields with global spinal structure in a multiscale CNN framework. For (2), the goal is to segment and identify the level of the vertebra in a textured 3D point cloud of the surgical field of view. Since 3D point clouds are unstructured (no regular grid) in nature, we will explore graphical convolutional networks for the task of semantic segmentation. The pre-operative and per-operative 3D spinal segmentations could then be used to track the spine during the surgery and to assist surgeons in the realignment of the spine.