fastMRI (the AI project for MRT)

2020: Start of software, the accelerating MRT-scanning by 4 times

In the middle of August, 2020 Facebook released in public access of software, the accelerating MRT-scanning by 4 times. The fastMRI project was carried out by research group on AI (FAIR) of Facebook and radiologists from Langone Health hospital of New York University. Together scientists trained model of machine learning at pictures of MRT of low and high resolution that the algorithm could recover pictures on the basis of 25% of the normal volume of input data. As a result when using this algorithm scanning can be executed quicker, significantly increasing satisfaction of patients.

The neural network understands the general structure of the medical image, - professor of a radiology of Langone Health Dan Sodickson explained. - It is possible to tell that the algorithm just fills unique features [images] of the specific patient on the basis of data retrieveds, using the previous knowledge as a basis.

Facebook provided software, the accelerating MRT-scanning by 4 times

Results of clinical trial were published in American Journal of Roentgenology together with the proofs confirming efficiency of a technique. During the research it was offered to radiologists to make the diagnosis on the basis of the traditional MRT-pictures and pictures recovered using AI. Doctors estimated pictures equally, having made the corresponding diagnoses. Thus, the traditional pictures and pictures recovered using AI were interchanged, and recovery on the basis of low amount of data did not lead to loss of important information and emergence of errors.

The fastMRI command notes that it was succeeded to avoid errors thanks to completeness of input data which cover the necessary area of a body entirely. Besides, scientists created the system of check of neural network on the basis of MRT-scanning physics. Thus, in the recovery time of pictures the AI system through regular periods checks whether there corresponds its output data to what can physically make the MRT-device.^[1]

2018: Release of base from 1.5 million pictures of MRT for training of medical AI

On November 30, 2018 department of a radiology of NYU School of Medicine in cooperation with Facebook AI Research (FAIR) opened the large-scale database of pictures of MRT open source. Originally the database will include more than 1.5 million anonymous MRT-pictures of a knee.

The base was published within fastMRI — the project which purpose is optimization of MRT-inspections using AI. The open database became a new phase of development of the project which Facebook and NYU announced in August, 2018. The fastMRI project will use AI for software development of MRT-devices that will allow to carry out scanning quicker, saving quality of images. The published database will help to train AI and will facilitate development of systems for MRT which will be able to work 10 times quicker than modern devices.

Facebook issued the database from 1.5 million pictures of MRT for training of medical artificial intelligence

The closed data access in the field of health care – one main obstacles in a way of artificial intelligence. Developers of the fastMRI project hope that their database will provide to researchers enough tools to overcome the problems facing them. The fastMRI database is the biggest collection of the unprocessed MRT-pictures open source. Future replenishments of base will include results of MRT-scanning of a liver and brain.

The fastMRI project managers consider that its successful implementation could reduce the need for anesthesia or sedative treatment at heavy patients, it is difficult to them to postpone long researches, and to provide universal access to MRT. Partnership of Facebook and NYU expands possibilities of research activity on AI and optimization of MRT. Researchers note that any discoveries made using fastMRI immediately will become property of scientific community.^[2]

Notes