fastMRI leverages adversarial learning to remove image artifacts

What it is:

As part of the fastMRI project to speed MRI scans by up to 10x, Facebook AI and NYU Langone Health researchers have developed a new way to use deep learning to address the problem of image artifacts in AI-accelerated MRIs. This technique, known as an orientation adversary, also significantly improves overall image quality.

The fastMRI team evaluated their method through a blinded study with six board-certified radiologists from NYU Langone Health. The results overwhelmingly showed that the orientation-adversarially trained model produced images with fewer artifacts and with no reduction in detail.

"The quality of our images is critically important in order to ensure the fidelity of signal abnormalities that form the basis of our diagnoses,” says Mitch Kline, one of the radiologists who participated in the study. “The orientation adversary reduced banding artifacts while preserving optimal signal to noise and image resolution."

We're sharing a large-scale dataset of MRI measurements and clinical images here.

How it works:

One challenge in using deep learning to generate highly accurate MRI scans from less raw data is that these reconstructions often suffer from banding and streaking artifacts which distract from or occlude image details. These artifacts may not be easy for nonexperts to see, but they are apparent to radiologists who are trained to recognize even the subtlest changes in the images.

"During our review of the accelerated images we noticed a significant horizontal banding artifact that significantly degraded image quality and had the potential to obscure pathology,” says Michael P. Recht, M.D., Chair and the Louis Marx Professor of Radiology at NYU Langone Health.

To address these issues, the fastMRI team leveraged adversarial training techniques to produce a deep learning model that takes raw data from accelerated MRI scans and produces accurate MRI images free of these artifacts. In adversarial learning, the training objective is augmented with an additional loss term that encourages the model to “trick” the adversary in some way. In this case, the fastMRI team used an adversary whose goal was to predict the orientation of the banding pattern. Banding with both horizontal and vertical patterns were produced during training by randomly transposing the input data both before and after reconstruction. The adversary and the reconstruction model were trained simultaneously, so that the adversary constantly adapted as the reconstructions improved until no banding remained.

Why it matters:

FastMRI is an open source, collaborative project that seeks to use AI to accelerate scans by 10x, which will make MRIs available to more people, shorten wait times, and reduce suffering for patients who find it difficult or impossible to stay inside the scanners for extended periods of time. To help the broader research community contribute to fastMRI and explore different approaches, we recently released a neuro MRI dataset and organized the first fastMRI Challenge .

Since image artifacts have been a major challenge with AI-accelerated MRIs, this new technique has the potential to move the project closer to implementation in clinical settings. Our technique is broadly applicable, as it may be used with any reconstruction model and dataset where fully sampled ground truth data is available.

Additionally, while state-of-the-art facilities today use 3 Tesla MRI machines, scanners with lower-strength magnets (1.5 Tesla, for example) are still commonly used around the world. These scanners tend to produce images with more banding artifacts. Our orientation adversary technique can help produce better reconstructions and accelerate scans produced with these devices, thereby delivering the benefits of accelerated MRIs to more people.

Read the paper:

https://arxiv.org/abs/2001.08699