MRI scans that are ten times faster with deep learning

Deep learning is of interest to the developers of the algorithms that are essential to the reconstruction of images produced through magnetic resonance. In 2020, a team from the CEA distinguished itself in this field.

In terms of magnetic resonance imaging (MRI), reducing examination time whilst maintaining excellent image quality is a real challenge. A research team from the CEA (the French Alternative Energies and Atomic Energy Commission) met this challenge by taking second place in the Brain fastMRI 2020 Challenge international competition. This is a collaborative research project between Facebook AI Research (FAIR) and NYO Langone Health, that aims to study the use of artificial intelligence (AI) in order to make IRM scans up to ten times faster.

The benefit of our method is that it is much higher performing than traditional algorithms when data is incomplete.

“Within NeuroSpin [NeuroSpin: a research center for neuroimaging with high field MRI at the Frédéric Joliot Institute for Life Sciences of CEA Paris-Saclay], my work consists in developing methods of MRI image acquisition acceleration on the one hand and improving the process of image reconstruction from the raw data collected by the MRI scanner on the other. It is precisely in this area that our work submitted to the Brain fastMRI 2020 Challenge was carried out”, explains Philippe Ciuciu, CEA Research Director at NeuroSpin and co-head of the joint Inria-CEA unit MIND [MIND: Models and Inference for Neuroimaging Data].

Competing with eighteen other research teams, the CEA team presented a new method of artificial intelligence based on neural networks. “Our model starts with the raw data and iteratively alternates the stages of improvement in the image space and of compatibility with the initial data. This is what we call a rolled-out system of neural networks”, Philippe Ciuciu specifies. “To this we have added a memory process between the different iterations and of all the stages, as these considerably improve the image quality and reconstruction time.”

A higher performing technology

The CEA team was ranked third during the competition’s first selection phase, which measured image quality (similarity in structure and signal-to-noise ratio) compared to reference data. The works of the three finalists were then subjected to a double-blind analysis by a group of six neuroradiologists. At the end of this final stage, the French researchers had moved up to second place.

“Artificial neural networks make it possible to reconstitute images of similar quality to that obtained from complete datasets but using only a small fraction of the data. Therefore, the benefit of our method is that it is much higher performing than traditional algorithms when data are incomplete, in this case when we only have access to an eighth of the data rather than a quarter”, Philippe Ciuciu highlights.

“This significant advantage enables considerable time savings as acquisition time is directly proportional to the quantity of data acquired. What’s more, for a 2D cut, our neural network only needs a tenth of a second to reconstitute an image, compared to one second for traditional algorithms, and our AI architecture is deployable on both 7 Tesla MRIs [Tesla: unit of measure of magnetic fields] and the 3 Tesla systems currently used in hospitals.

Benefits for both patients and healthcare facilities

What deep learning brings in terms of image acquisition time reduction without reducing image quality provides a glimpse of the many benefits for both patients and healthcare facilities.

The possibility to spend less time in MRI machines to undergo scans is of substantial comfort. Image quality is also improved as the artifacts due to patients motions during scans are de facto also reduced. Finally, the reduction in scan time for each patient makes it possible to increase the number of patients that a hospital can treat per day.

Since its presentation in 2020 at the Brain fastMRI Challenge, the technology developed by the CEA team has not stopped improving. “We have extended our neural network architecture to cover more ambitious and more complex acquisition scenarios, either in 3D (resolution gain), or in non-Cartesian imaging (shorter examinations), by offering better correction of artifacts”, Philippe Ciuciu explains.

Last April, the researcher also submitted an ERC (European Research Council) Advanced application, with financing of 3 million euros over five years, to develop the CEA system in the scope of earlier and better differentiated diagnosis of people with Parkinson’s disease.

“Furthermore, in a clinical practice approach, the Paris Brain Institute has included us in the establishment of a hospital-university network aimed at demonstrating the efficiency of the 7 Tesla MRI in the care of illnesses such as Parkinson’s disease, epilepsy, and brain tumors.” The funding decisions for the ERC and for the Brain Institute should fall respectively in March and September 2023.

Sources :

Read also on Hello Future

Healthcare: will algorithms soon be taking decisions instead of doctors?


Generative AI helps to design new drug candidates


Is AutoML the key to the democratization of machine learning?


AI: BigScience brings transparency and variety to large language models


Using AI to Predict Diseases in Olive Groves

A woman talks to her smartphone

Never Too Many Neurons in Speech Recognition


When AI knows your state of health thanks to the sound of your voice


Video Compression: A Neural Network Challenging the Conventional Approach