Abstract
Purpose: In diffusion MRI (dMRI), the volumetric and bundle analyses ofwhole-brain tissue microstructure and connectivity can be severely impeded byan incomplete field-of-view (FOV). This work aims to develop a method forimputing the missing slices directly from existing dMRI scans with anincomplete FOV. We hypothesize that the imputed image with complete FOV canimprove the whole-brain tractography for corrupted data with incomplete FOV.Therefore, our approach provides a desirable alternative to discarding thevaluable dMRI data, enabling subsequent tractography analyses that wouldotherwise be challenging or unattainable with corrupted data. Approach: Wepropose a framework based on a deep generative model that estimates the absentbrain regions in dMRI scans with incomplete FOV. The model is capable oflearning both the diffusion characteristics in diffusion-weighted images (DWI)and the anatomical features evident in the corresponding structural images forefficiently imputing missing slices of DWI outside of incomplete FOV. Results:For evaluating the imputed slices, on the WRAP dataset the proposed frameworkachieved PSNRb0=22.397, SSIMb0=0.905, PSNRb1300=22.479, SSIMb1300=0.893; on theNACC dataset it achieved PSNRb0=21.304, SSIMb0=0.892, PSNRb1300=21.599,SSIMb1300= 0.877. The proposed framework improved the tractography accuracy, asdemonstrated by an increased average Dice score for 72 tracts (p < 0.001) onboth the WRAP and NACC datasets. Conclusions: Results suggest that the proposedframework achieved sufficient imputation performance in dMRI data withincomplete FOV for improving whole-brain tractography, thereby repairing thecorrupted data. Our approach achieved more accurate whole-brain tractographyresults with extended and complete FOV and reduced the uncertainty whenanalyzing bundles associated with Alzheimer's Disease.