With magnetic resonance imaging, it is possible to "see" the brain in action. The most widely used method is functional BOLD MRI (fMRI BOLD).  However, this method is an indirect indicator of neuronal activity: it yields spatial maps reflecting local and transient variations in blood flow that accompany neuronal activation, a phenomenon known as the neurovascular coupling. This coupling is usually considered a linear time-invariant system and is represented by a mathematical function, the hemodynamic response function (HRF). By the very nature of its significance, the HRF varies from one region of the brain to another, from one individual to another, from one state (normal/pathological) to another, etc. Estimating it individually (by a mathematical deconvolution operation), both in each individual and within each brain area, remains challenging as it consists in solving a large scale inverse problem. However this challenge is worthful as the outcome yields local time-resolved neural activities from BOLD fMRI signals. 

Several approaches are already used to estimate the HRF in fMRI task paradigm (activation fMRI). In that cases, these methods fit a model to explain the BOLD signal from the experimental paradigm, i.e., from the expected brain activity based on the experiment the subject is engaged in. One of the limitations is that the experimental paradigm is a kind of substitute for real brain activity and does not account for possible differences in response time between subjects. Moreover, it is not possible to use such methods to estimate HRF in fMRI without experimental paradigm (resting-state fMRI).

Researchers from the PARIETAL team (NeuroSpin), in collaboration with the BioMaps unit (SHFJ) and the University of Edinburgh, propose a deconvolution method that simultaneously identifies regional hemodynamic coupling, characteristic signals of neural activity and the associated brain regions. This method is versatile and can process resting or activation fMRI data, in a very short time (1 min per brain) thanks to an optimized implementation. It is available as an open-source Python module (hemolearn: https://github.com/hcherkaoui/hemolearn)

Several validations have been conducted. At the scale of an individual, the researchers were able to identify the main functional brain networks at rest and to access their underlying neural dynamics. On the scale of a group (48 individuals) from the UK Biobank cohort, this method made it possible to quantify an index of interhemispheric hemodynamic variability in order to automatically discriminate patients who had suffered a stroke from healthy subjects. Similarly, they showed on 459 individuals divided between seniors (64 to 70 years old) and younger (40 to 44 years old), that a prolonged hemodynamic delay is a good predictor (75% accuracy) of cerebral aging.