This Artificial Intelligence (AI) Research From NYU and Meta AI Proposes A Deep Learning Technique That Can Reconstruct Missing Data From Rapid MRI Scans

A recent study from the NYU Grossman School of Medicine and Meta AI Research shows that artificial intelligence (AI) can rebuild coarse-sampled, fast MRI scans into high-quality pictures with equal diagnostic value as those created using standard MRI. According to the study, MRI images may be made available to more patients, and appointment wait times might be cut in half if they were reconstructed using AI rather than traditional methods. Researchers from Meta AI and NYU Langone’s imaging experts and radiologists collaborated on an AI model to speed up MRI. It also generated the world’s biggest raw MRI data repository, which researchers and developers have utilized in various fields.

The NYU Langone scientists eliminated nearly three-quarters of the raw data gathered by traditional, sluggish MRI scans to replicate expedited scans in a previous “proof-of-principle” study. Faster MRI scans were used to train an artificial intelligence model that produced pictures that were otherwise indistinguishable from those produced by slower scans. Similar to how the brain constructs pictures by filling in missing visual information from the local context and past experiences, the researchers in this new study performed expedited scans with just one-fourth of the real data and used the AI model to “fill in” the missing information. The fastMRI scans were accurate and of higher quality than the conventional scans in both experiments.

We used 298 clinical 3-T knee assessment pictures to train a DL reconstruction model. Between January 2020 and February 2021, patients who were clinically referred for knee MRI completed a conventional accelerated knee MRI protocol at 3 T followed by an accelerated DL procedure for a prospective study. It was determined whether or not the DL reconstructed pictures were interchangeable with the traditional images in anomaly identification. Six musculoskeletal radiologists looked at each exam. Ordinal ratings from 0 to 4 were used in analyses evaluating the probability of anomalies in meniscal or ligament tears, bone marrow, and cartilage. Overall image quality, presence of artifacts, sharpness, and signal-to-noise ratio were all evaluated, and four-point ordinal values were used to compare the methods.

170 people were assessed (mean age SD: 45 16; 76 male). It was found that the DL-reconstructed photos were just as good as the traditional images in spotting anomalies. On average, DL photos received a higher quality rating among six readers than traditional photographs (P .001). The radiologists agreed that the AI-reconstructed images were just as good as the traditional images for making diagnoses of tears and abnormalities. They also decided that the faster scans had far higher image quality overall.

Researchers stress that no unique tools are needed to perform FastMRI. Standard MRI machines may be programmed to collect fewer data than is often required. The fastMRI effort has released its data, models, and code as an open-source project for use by other researchers and makers of commercial MRI systems.

With fastMRI, the time it takes to perform an MRI scan, which may take up to 30 minutes, is reduced to less than 5 minutes, bringing it on par with the time it takes to perform an X-ray or CT scan. In contrast to these other imaging modalities, however, MRI offers a wealth of information, such as seeing soft tissues from numerous angles, highlighting microscopic cartilage anomalies, and pinpointing abdominal malignancies.

In conclusion, deep learning reconstruction of prospectively accelerated knee MRI allowed for a near-twofold decrease in scan time, enhanced picture quality, and had equal diagnostic usefulness compared to traditional reconstruction. It has been shown that deep learning reconstruction may cut scan time for a knee MRI by about half compared to the standard procedure without sacrificing diagnostic accuracy.

Check out the Paper and Reference Article. All Credit For This Research Goes To the Researchers on This Project. Also, don’t forget to join our Reddit Page, Discord Channel, and Email Newsletter, where we share the latest AI research news, cool AI projects, and more.