Unsupervised Normative Learning for Quality Assessment on Diffusion MRI

Quality assessment of diffusion MRI (dMRI) aims to ensure that the presence of image artifacts does not affect the results of subsequent image analysis. Recently, a variety of deep learning methods have been proposed to enhance the quality and utility of these acquisitions. Still, they require a large amount of training data with artifact labeling. In this work, we developed an Unsupervised Quality Assessment tool (UNL-QA) for dMRI data that can handle a variety of artifacts such as ghosting, spike, noise and swap. Our method uses a vector quantized variational autoencoder to train typical patterns on a standard dMRI dataset, achieving 99.732% accuracy in detecting artifacts and standard data and 95% sensitivity in detecting artifacts on the training dataset. Performance across datasets is also excellent. It has also been demonstrated to be replicable on other datasets with different acquisition parameters. Without labeled artifact images necessary, UNL-QA allows fast and paves the way for efficient and effective artifact detection in large datasets. The code is available at https://github.com/yjh200319/UNL-QA.git.