NeuroSpin’s organization includes four research units of which the Inria-CEA MIND and the CEA-CNRS BAOBAB labs play pivotal roles in the use of AI, signal processing and MR physics for developing new accelerated MR imaging techniques, notably for functional imaging (fMRI).
“EXPLORE +” BlueSky project is a large scale CEA-funded project (2M€ over 4 years since 2023) awarded to Dr Philippe Ciuciu. The project aims to understand learning and decision making processes in the healthy brain at the mesoscale using fMRI imaging at high spatiotemporal resolution (e.g. up to 500μm and < 1s). This cognitive neuroscience study is supervised by Dr Florent Meyniel’s team at NeuroSpin.
The postdoc fellow will contribute in collaboration to a recently hired PhD candidate to making this extremely challenging fMRI acquisition scenario realistic both on the fMRI data acquisition and image reconstruction sides. The PhD candidate will use and extend 3D SPARKLING, a non-Cartesian readout (1-4), which embodies strong acceleration potential to enable high spatial and temporal resolution fMRI without impeding whole brain coverage. Additionally, SPARKLING acquisitions currently require a computationally-demanding image reconstruction process prior to conducting statistical analysis for the detection of evoked brain activity during the EXPLORE+ experimental paradigm.
Consequently, the postdoc fellow that will be recruited will work on improving the existing fMRI image reconstruction pipeline to make it easier and faster for end-users, i.e. cognitive neuroscientists. For doing so, he/she will rely first on in-house developments notably the PySAP software package and its plugin for fMRI. Second, he/she will investigate deep learning solutions based on unrolled deep neural networks (5) and/or Plug&Play algorithms (6). When addressing deep learning solution for image reconstruction in fMRI, as ground truth, i.e. fully sampled data cannot be collected first due to the short scan time constraint and second because brain activity is never completely reproducible, two competing solutions can be envisaged: either training DL architectures on realistic synthetic data, such as those yielded by our recent realistic SNAKE-fMRI simulator (7), and then proceeding to transfer learning or domain adaptation, or using self-supervised approaches (8). Both methods will be investigated. Ablation studies will be conducted on simulation and real fMRI data to demonstrate the robustness and added-value of the proposed approach compared to compressed sensing solutions.
A very important deliverable will be to implement a complete reconstruction solution directly connected at the console using “Gadgetron” (9) or “Open Recon” Siemens Healthineers platform (10). Automating the correction of off-resonance artifacts within the pipeline is also a key aspect.