2023
View Abstract / View Presentation
Super-resolution and CNN denoising to improve the accuracy of small brainstem structure characterization with in vivo diffusion MRI
Benjamin Ades-Aron1, Hong-Hsi Lee1, Heidi Schambra2, Dmitry S. Novikov1, Els Fieremans1, and Timothy Shepherd1
1Radiology, NYU School of Medicine, New York, NY, United States, 2Neurology, NYU Langone, New York, NY, United States
We describe and evaluate a novel combination of Fast Gray Matter Acquisition T1 Inversion Recovery (FGATIR) denoising with deep learning and super-resolution techniques to improve the accuracy of small internal brainstem structure segmentation on advanced diffusion MRI data
Co-registered axial images of mid-pons using FGATIR, standard and super-resolution fractional anisotropy with overlaid segmentations of corticospinal tract (PON; dark yellow), pontine reticular formation (PRF; light yellow) and medial longitudinal fasciculus (MLF; red) created by neuroanatomy expert using FGATIR contrast propagated to the diffusion data. The round shape of the PON segmentation is better preserved with super-resolution. There is less CSF contamination of the MLF from the 4th ventricle with super-resolution.
Overview of the pipeline. Images are labelled as solid blue blocks and processing steps are labelled as light blue blocks. The pipeline relies on denoised FGATIR data that undergoes diffeomorphic registration through a 3D T2 weighted image to a super-resolution enhanced diffusion weighted dataset.