MRI2Mesh: Intervertebral Disc Mesh Generation From Low Resolution MRI Using Graph Neural Networks with Cross Level Feature Fusion

This paper tackles the challenge of reconstructing high resolu- tion meshes of intervertebral discs from low resolution thick- sliced sagittal MR images of the lumbar spine. By generating the meshes directly from the volume, we eliminate the error propagation incurred when segmentation is used as an inter- mediary step. Our method is based on a combination of Con- volutional and Graph Neural Networks. We propose a novel cross-level disc feature fusion scheme, addressing both local and global shape context. This differs fundamentally from other surface reconstruction approaches, since our method ex- ploits the underlying anatomical characteristics via an axial attention transformer. Experimental results show that the pro- posed method improves the Hausdorff distance by 5.87% and point-to-surface distance by 14.5% when compared to other baseline models.