A team from the ITACA-UPV has led the development of an artificial intelligence method capable of synthesising high-quality magnetic resonance imaging (MRI) scans without the need to acquire all the usual sequences. This new method will allow MRI examinations to be faster, more comfortable and more cost-effective, helping to improve the detection of neurological diseases.
The team’s next challenge is to extend the technique to other sequences such as FLAIR (Fluid Attenuated Inversion Recovery), a variant of T2-weighted images that suppresses the cerebrospinal fluid signal and shows lesions associated with diseases such as Alzheimer’s, multiple sclerosis or brain tumours.
The work, published in Imaging Neuroscience, was coordinated by the Medical Image Analysis Group (MIA-LAB) at the ITACA Institute of the UPV. Researchers from the Applied Mathematics Department at UPV, the Psychobiology Department at the University of Valencia, the Medical Imaging Area at La Fe University and Polytechnic Hospital, the Joint Biomedical Imaging Unit (UMIB-IA) at Fisabio-CIPF, the French National Centre for Scientific Research and the University of Bordeaux also contributed.
Deep neural network in 3D
This new method is based on a 3D deep neural network that generates T2-weighted images — highly sensitive to the presence of water, allowing the detection of oedema, inflammation or ischaemia — from T1-weighted images, which provide a detailed anatomical representation of the brain and sharply distinguish white matter from grey matter. In this way, T1 images provide the “structure”, while T2 and FLAIR highlight potential pathological changes.
To achieve this, the system integrates prior anatomical information and uses semi-supervised learning techniques, an artificial intelligence approach that combines a small number of medically labelled images with a large volume of unlabelled data. This enables the training of powerful models without requiring fully annotated datasets.
“In an MRI examination, each type of image provides different information about the brain, but acquiring them all makes the procedure longer, more expensive and sometimes uncomfortable. Our system can generate the missing images from those already acquired, reducing time and resources,” explains Sergio Morell, main author of the study.
Innovation and international validation
The method led by the UPV researchers combines real anatomical knowledge, specific training strategies and a semi-supervised approach that improves its ability to generalise across patients and scanners. In brain segmentation tests, it outperformed the most advanced available techniques, even in complex cases such as brains with lesions or high anatomical variability. It also produces results within seconds, enabling its application in clinical settings.
The study was funded by the Spanish Ministry of Science, Innovation and Universities and the French National Research Agency.
REFERENCE: Sergio Morell-Ortega, Marina Ruiz-Perez, Marien Gadea, Roberto Vivo-Hernando, Gregorio Rubio, Fernando Aparici, Mariam de la Iglesia-Vaya, Thomas Tourdias, Boris Mansencal, Pierrick Coupé, José V. Manjón; Robust deep MRI contrast synthesis using a prior-based and task-oriented 3D network. Imaging Neuroscience 2025; 3 IMAG.a.116. doi: https://doi.org/10.1162/IMAG.a.116
