Inpainting Cropped Diffusion MRI using Deep Generative Models.

Area of Science:

• Medical Imaging
• Neuroscience
• Artificial Intelligence

Background:

• Image acquisition artifacts, like cropped fields of view in MRI, can lead to data loss and reduced analysis power.
• Variational autoencoders (VAE) are deep generative models effective for high-resolution image reconstruction.

Purpose of the Study:

• To evaluate deep learning models for restoring missing head regions in MRI scans.
• To improve the quality and reduce noise in diffusion-weighted imaging (DWI) processing.

Main Methods:

• Utilized diffusion-weighted images (DWI) from the National Consortium on Alcohol and Neurodevelopment in Adolescence (NCANDA) dataset.
• Compared U-Net, VQVAE, VAE-GAN, and a novel U-VQVAE model for inpainting missing cranial regions.
• Assessed accuracy of inpainting and its impact on fractional anisotropy (FA) in the supplementary motor area.

Main Results:

• The proposed U-VQVAE model achieved the highest accuracy in restoring missing head regions.
• Inpainting with U-VQVAE resulted in lower fractional anisotropy (FA) in the supplementary motor area compared to original MRIs.
• This indicates reduced processing noise and enhanced quality of generated results.

Conclusions:

• The U-VQVAE model effectively restores missing head regions in MRI scans, mitigating data loss from cropping artifacts.
• This deep learning approach enhances the quality of DWI processing, leading to more reliable neuroimaging analyses.
• The developed method offers a solution for improving the robustness of neuroimaging studies using MRI data.