Improving quality of spinal cord DTI using inpainting

The goal of this multi-disciplinary proposal is to build upon the existing collaboration between TUHH and UKE with the aim to refine inpainting of diffusion tensor imaging (DTI) data such that it reliably and automatically identifies artefactual voxels in clinical spinal cord DTI data, and re- places them with corrected values. An important step in the identification process is an adequate image segmentation mechanism that can improve the detection of artefacts.

First, the error of the diffusion tensor-model fit will be used to automatically identify artefactual data points (an approach that has been previously used in outlier-rejection methods). To improve the segmentation of artefactual data points we will use convolutional neural networks and deep learning. Then, the inpainting algorithm will be used to replace the artificial voxels using information from the adjacent voxels, instead of removing these data points, thereby increasing the noise level (as done in outlier-rejection).

The proposed algorithm was implemented and validated using simulation data. On the basis of this project, a new collaborative research project between the UKE and TUHH was created: automated segmentation of 2D and 3D objects using analytical and deep-learning methods, including one publication (Tabarin et al., 2019), one internship project (Najafi), a bachelor (B.Sc: Klisch), a master thesis (M.Sc: Przybyla) and two PhD projects (PhD: Ashtarayeh and Mordhorst). Part of these developments was the consideration of geometric deep learning, a strategy that extends convolutional neural networks to data defined on graphs (Klisch and Przybyla). This step is an additional component for an adequate detection of artefacts in more complex settings. Moreover, two studies, one on spinal cord imaging (David et al., 2019) and one on artefact removal (Papazo- glou et al., 2019), were published during the funding period of this project.