An Unsupervised Deep Learning Method for CT and MR Registration in Spine Surgery Simulator

Image registration is a fundamental step in patient specific spine surgery simulation where two images are aligned in the same spatial geometry coordinate space. Registration of Computed Tomography (CT) and Magnetic resonance imaging (MRI) has important implications for 3D model creation, clinical diagnosis, treatment planning, and image-guided surgery as it provides complimentary information obtained from different image modalities. Existing methods are slow and requires manual intervention. Therefore, a more accurate, robust and fast method was introduced with the implementation of a VoxelMorph framework using Modality Independent Neighbourhood Descriptor (MIND) based loss function. A UNET based model was trained with 268 pairs of CT and MRI images acquired through Sunnybrook Health Science Centre. The trained model was tested through 10 pairs of test data with vertebral body segmentation. The results achieved good performance accuracy; Dice Similarity Coefficient (DSC) : 0.845 ± 0.028 and Hausdorff distance : 1.128 ± 0.588 mm with regularization parameter (λ) tuned to 0.7. Integration of this method into a spine surgery simulation workflow will allow a fast (few seconds in GPU, under a minute in CPU), accurate and robust registration process.