NYU Gross School of Medicine and Meta AI Research found that AI can generate MRI scans faster than traditional methods. As a result, it could lead to more accessibility and lower wait times.

Traditional MRI scans have many advantages, but they also have some drawbacks, such as the extended wait times patients may have to endure.

Objective

The fastMRI project, launched in 2018 by NYU Langone Health and Meta AI Research, sought to employ artificial intelligence to speed up MRI scans. In a previous proof-of-concept study, the NYU Langone group eliminated 75% of raw data from regular MRI scans to simulate faster scans. As a result, images from the fastMRI scans were just as high-quality as those from the traditional scans.

For the latest study, researchers utilised expedited scans using only 75% of the data and let the AI model complete the remainder. One hundred and seventy people took part in the trial, and each had a diagnostic MRI of their knee performed using either a standard MRI or a sped-up AI protocol.

Six musculoskeletal radiologists analysed the MRI scans to detect meniscal or ligament tears and bone marrow or cartilage abnormalities. To reduce the potential for bias, they were not informed which photographs had been digitally recreated using AI. However, after testing, they decided the AI-reconstructed photos were on par with gold-standard images for spotting tears or anomalies. In addition, accelerated scans had better image quality than regular scans.

fastMRI's efficiency and open-source nature have been lauded. The researchers also pointed out that fastMRI can complete exams in a similar amount of time to X-rays or CT scans but with better results.

A similar computer vision model that expedited and enhanced the quality of head MRI examinations was the subject of a March 2022 study. The researchers concentrated their attention on axial T2-weighted scans since they made up more than 90% of MRIs. They developed computer vision models through labelling datasets, image preprocessing, model interpretability, experimentation, and simulation analysis.

Furthermore, the best-performing model, as discovered by the research team, has the potential to facilitate the generation of MRIs that are both quicker and more accurate.

Impact

AI-based technologies provide precise volumetric measurements to evaluate potential brain tissue preservation after injury and automatically fill this information into dictation platforms and other systems. Recent research has shown that with the help of AI, reconstructing MRI scans may be done much more quickly and accurately than before.

Conclusion

Deep learning reconstruction of prospectively accelerated knee MRI reduced scan time by nearly two-thirds, resulted in higher-quality images, and was just as useful for diagnosis as traditional reconstruction. However, despite its diagnostic capacity, MRI requires considerable time to acquire. Deep learning (DL) has lately enabled quick, high-quality picture reconstructions from undersampled data, although its utility in clinical practice still needs to be determined. This study compares conventionally and prospectively accelerated DL-reconstructed MRI for internal knee derangement.

Furthermore, an MRI of the knee took roughly half as long to acquire using deep learning reconstruction as it did using the standard methodology without sacrificing diagnostic accuracy.