Diffusion MRI Fibre Orientation Distribution Inpainting

Zihao Tang, Xinyi Wang, Mariano Cabezas, Arkiev D'Souza, Fernando Calamante, Dongnan Liu, Michael Barnett, Chenyu Wang, Tom Weidong Cai

Abstract: The analysis of diffusion weighted brain magnetic resonance images, including the estimation of fibre orientation distribution (FOD), tractography, and connectomics, is a powerful tool for neuroscience research and clinical applications. However, focal brain pathology and imaging acquisition artifacts affecting white matter tracts may disrupt or corrupt FOD values respectively, invalidating tractography and connectome reconstructions. In this work, we propose a 3D FOD inpainting framework, named order-wise coefficient estimation network (OCE-Net), to dynamically reconstruct the affected regions. Our feature encoding stage, based on gated convolutions, extracts features from all the input FOD coefficients and re-weights them using channel attention and independent order-wise decoders, to independently predict the coefficients for each spherical harmonic order. We evaluated our model on a subset of scans from the HCP dataset, and conducted tractography and connectomics to further analyse the impact of inpainting. Our experimental results, including a statistical analysis of the reconstructed connectomes, show that our OCE-Net approach can successfully reconstruct the original FODs in the focally disrupted regions.