Bridging Cartesian and non-Cartesian sampling in MRI

MRI is a widely used neuroimaging technique used to probe brain tissues, their structure and provide diagnostic insights on the functional organization as well as the layout of brain vessels. However, MRI relies on an inherently slow imaging process. Reducing acquisition time has been a major challenge in high-resolution MRI and has been successfully addressed by Compressed Sensing (CS) theory. However, most of the Fourier encoding schemes under-sample existing k-space trajectories which unfortunately will never adequately encode all the information necessary. Recently, the Mind team has addressed this crucial issue by proposing the Spreading Projection Algorithm for Rapid K-space sampLING (SPARKLING) for 2D/3D non-Cartesian T2* and susceptibility weighted imaging (SWI) at 3 and 7 Tesla (T) 130, 131, 6.

These advancements have interesting applications in cognitive and clinical neuroscience. However, the original SPARKLING trajectories are prone to off-resonance effects due to susceptibility artifacts. Therefore, in 2023, two years ago, we extended the original SPARKLING methodology along two diretions : and developed the MORE-SPARKLING (Minimized Off-Resonance Effects) approach to correct for these artifacts. The fondamental idea implemented in MORE-SPARKLING is to make different k-space trajectories more homogeneous in time, in the sense that samples supported by different trajectories that are close in k-space must be collected at approximately the same time point. This allows us to mitigate the issue of off-resonance effects (signal void, geometric distortions) without increasing the scan time, as MORE-SPARKLING trajectories have exactly the same duration as their SPARKLING ancestor. This approach was published 5. Additionally, we also extended SPARKLING in another direction, namely the way we sample the center of k-space and proposed the GoLF-SPARKLING version in the same paper 5. The core idea was to reduce the over-sampling of the center of k-space and grid it to collect Cartesian data and make notably the estimation of sensitivity maps in multicoil acquisition easier.

In 2024, we have merged GoLF-SPARKLING with well-established clinically used parallel imaging techniques like GRAPPA and CAIPIRINHA. The latter are mostly limited to Cartesian sampling, and their extension to non-Cartesian sampling is not direct.In 52 we proposed to extend the SPARKLING frameworkto ensure GRAPPA accelerated Cartesian sampling in the k-space center while allowing non-Cartesian sampling in the periphery. We can now enforce Cartesian acceleration in the center of k-space through affine constraints. Using $2\times 2$ GRAPPA with SPARKLING trajectories for anatomical MRI with MPRAGE sequences enables 10x acceleration, achieving 1mm isotropic whole-brain scans in 1 minute (see Fig. 1).

SPARKLING for fMRI

Additionally, we have shown that 3D-SPARKLING is a viable imaging technique and good alternative to Echo Planar Imaging for resting-state and task-based fMRI 11 and 12. This is illustrated in Fig. 2 during a retinotopic mapping experiment which consists in mapping the retina to the primary visual cortex. These results have been obtained at a 1mm isotropic resolution both for EPI and SPARKLING acquisitions.

Deep Learning for 3D Non-Cartesian MR Image reconstruction

Deep learning and notably unrolled neural netowrk architectures have shown great promise for MRI reconstruction from undersampled data. However, there is a lack of research on validating their performance in the 3D parallel imaging acquisitions with non-Cartesian undersampling. In addition, the artifacts and the resulting image quality depend on the under-sampling pattern. To address this uncharted territory, in 2024 we extended the Non-Cartesian Primal-Dual Network (NC-PDNet) 160, to a 3D multi-coil acquisition setting. We evaluated the impact of channel-specific versus channel-agnostic training configurations and examined the effect of coil compression. Finally, using the publicly available Calgary-Campinas dataset, we benchmarked four distinct non-Cartesian undersampling patterns, with an acceleration factor of six. Our results in Fig. 3 showed that NC-PDNet trained on compressed data with varying input channel numbers achieves an average PSNR of 42.98dB for 1 mm isotropic 3- channel whole-brain 3D reconstruction. With an inference time of 4.95sec and a GPU memory usage of 5.49 GB, our approach demonstrates significant potential for clinical research application.

Fast reconstructions of ultra-high resolution MR data from the 11.7T Iseult scanner

Open-source MR reconstruction tools often fail to efficiently utilize GPU resources and lack support for generalized GRAPPA implementations. Many tools are limited to 2D or 3D reconstruction, and few incorporate advanced techniques such as 2D-CAIPIRINHA, which enhances imaging capabilities. Our new open source gGRAPPA GPU accelerated Python package aims to provide a fast, flexible, open-source tool for generalized GRAPPA/CAIPI reconstruction  45. Using PyTorch, gGRAPPA runs multiple convolutional windows in batch mode to optimize GPU memory usage and accelerate reconstruction times. gGRAPPA achieves up to a 65x speedup over CPU implementations and a 6x speedup compared to non-batched GPU methods, enabling efficient and fast reconstruction of MRI scans. This tool has allowed us to outperform the Siemens image reconstructor, in terms of speed, on the world premiere 11.7T MR system (Iseult scanner available at NeuroSpin) to reconstruct the first in vivo $T_2^{*}$ weighted MR images at a 200-µm in plane resolution.

SNAKE-fMRI: A realistic fMRI data simulator for high resolution functional imaging

Functional Magnetic Resonance Imaging (fMRI) has emerged as a powerful non-invasive tool in neuroscience, enabling neuroscientists to understand human brain functions. However, fMRI acquisition and image reconstruction techniques are complex to optimize and benchmark due to the lack of ground truth that produces absolute and quantitative metrics. Repeating in-vivo experiments may face the issue of limited reproducibility, which is time-consuming and expensive. fMRI simulators have been developed to generate synthetic fMRI images, where brain responses are artificially added to existing or artificial data. However, they don’t simulate the complete MR acquisition process, lack flexibility, integration with post-processing, and computational efficiency. Exploring new acceleration schemes in the acquisition setting and innovative reconstruction methods with these tools is not feasible. To address these unmet needs, in 20204 we have developed SNAKE-fMRI 51, an open-source fMRI simulator that operates in both image and k-space domains to yield realistic synthetic fMRI data. Its flexibility allows us to investigate various acquisition setups regarding SNR and acceleration factors to reach higher spatial and temporal resolution and validate reconstruction methods against those scenarios. Its key principles are summarized in Fig. 5 and some comparative results are shown in Fig. 6.

The interest of this simulator will be to provide ground truth data to train deep learning models for fMRI reconstruction in 2025.

NeuroConText: Contrastive text-to-brain mapping for neuroscientific literature

Hundreds of neuroscientific articles are published annually, highlighting the continuous growth and expansion of knowledge in this field. These contributions from scientists and researchers provide new insights and findings that enhance our understanding of brain functions. Meta-analysis is a statistical tool that combines results from multiple studies to improve the reliability and generalizability of neuroscientific findings. It serves three key purposes: First, it synthesizes information from the literature to build the state of the art by offering a consolidated view of what is currently known. This allows researchers to see where the field stands collectively and highlights consensus or inconsistencies. Second, meta-analysis provides context to interpret new results by comparing novel experimental data with existing patterns in the meta-analysis of the literature, which clarifies how new data align with or deviate from established findings. Third, meta-analysis generates hypotheses on candidate brain regions or relevant cognitive domains by revealing patterns that might be central to a phenomenon.

In 34, we introduced NeuroConText, a novel coordinate-based meta-analysis (CBMA) tool to bridge the three heterogeneous modalities commonly found in neuroscientific studies: text, reported brain activation coordinates, and brain images. NeuroConText uses neuroscientific articles to extract their text and activation coordinates. We benefited from the embeddings of advanced large language models (LLM) for text feature representation. Additionally, we used Kernel Density Estimation (KDE) to reconstruct brain maps from the coordinates  170, 169. To address the high dimensionality of the brain images reconstructed by KDE, we employed the dictionary of functional modes (DiFuMo), a probabilistic atlas that effectively reduces data dimensionality  99.

Then, NeuroConText defines a shared latent space between text and coordinates, using contrastive learning to retrieve the brain activation coordinates corresponding to the input text. NeuroConText can analyze long texts, leveraging the complete information in articles’ text to enhance the accuracy of text-to-brain associations. By incorporating advanced language models like Mistral-7B, it processes complex neuroscientific text and extracts the semantic in the text  125. NeuroConText considerably outperforms existing regression-based state-of-the-art methods NeuroQuery and Text2Brain in associating text with brain activations, achieving a threefold improvement in the retrieval task. To improve our understanding of the internal mechanisms of the NeuroConText model, we also evaluated its ability to reconstruct brain activation contrasts from text latent representations using descriptions of the NeuroVault dataset  115. The quality of these reconstructed activation maps is found to be comparable with state-of-the-art baselines.

Improved priors for inverse problem resolution

Selecting an appropriate prior to compensate for information loss due to the measurement operator is a fundamental challenge in inverse problems. Implicit priors based on denoising neural networks have become central to widely-used frameworks such as Plug-and-Play (PnP) algorithms. During this year, we made several progress to better design implicit priors based on existing denoisers. First, we proposed to enforce equivariance to certain groups of transformations (rotations, reflections, and/or translations) on the denoiser used as implicit priors. This simple procedure strongly improves the stability of the algorithm as well as its reconstruction quality. We derived theoretical insights which highlight the role of equivariance on better performance and stability, and experiments on multiple imaging modalities and denoising networks show numerically these benefits. Then, we introduce Fixed-points of Restoration (FiRe) priors as a new framework for expanding the notion of priors in PnP to general restoration models beyond traditional denoising models. The key insight behind FiRe is that natural images emerge as fixed points of the composition of a degradation operator with the corresponding restoration model. Adopting this fixed-point perspective, we show how various restoration networks can effectively serve as priors for solving inverse problems. Experimental results validate the effectiveness of FiRe across various inverse problems, establishing a new paradigm for incorporating pretrained restoration models into PnP-like algorithms.

Statistically Valid Variable Importance Assessment through Conditional Permutations

Variable importance assessment has become a crucial step in machine-learning applications when using complex learners, such as deep neural networks, on large-scale data. Removal-based importance assessment is currently the reference approach, particularly when statistical guarantees are sought to justify variable inclusion. It is often implemented with variable permutation schemes. On the flip side, these approaches risk misidentifying unimportant variables as important in the presence of correlations among covariates. Here we develop a systematic approach for studying Conditional Permutation Importance (CPI) that is model agnostic and computationally lean, as well as reusable benchmarks of state-of-the-art variable importance estimators. We show theoretically and empirically that CPI overcomes the limitations of standard permutation importance by providing accurate type-I error control. When used with a deep neural network, CPI consistently showed top accuracy across benchmarks. An experiment on real-world data analysis in a largescale medical dataset showed that CPI provides a more parsimonious selection of statistically significant variables. Our results suggest that CPI can be readily used as drop-in replacement for permutation-based methods.

False Discovery Proportion control for aggregated Knockoffs

Controlled variable selection is an important analytical step in various scientific fields, such as brain imaging or genomics. In these high-dimensional data settings, considering too many variables leads to poor models and high costs, hence the need for statistical guarantees on false positives. Knockoffs are a popular statistical tool for conditional variable selection in high dimension. However, they control for the expected proportion of false discoveries (FDR) and not their actual proportion (FDP). We present a new method, KOPI, that controls the proportion of false discoveries for Knockoff-based inference. The proposed method also relies on a new type of aggregation to address the undesirable randomness associated with classical Knockoff inference. We demonstrate FDP control and substantial power gains over existing Knockoff-based methods in various simulation settings and achieve good sensitivity/specificity tradeoffs on brain imaging and genomic data.

Clinical biomarkers for epilepsy

Postsurgical seizure freedom in drug-resistant epilepsy (DRE) patients varies from 30% to 80%, implying that in many cases the current approaches fail to fully map the epileptogenic zone (EZ). We aimed to advance a novel approach to better characterize epileptogenicity and investigate whether the EZ encompasses a broader epileptogenic network (EpiNet) beyond the seizure zone (SZ) that exhibits seizure activity. We first used computational modeling to test putative complex systems-driven and systems neuroscience-driven mechanistic biomarkers for epileptogenicity. We then used these biomarkers to extract features from resting-state stereo-electroencephalograms recorded from DRE patients and trained supervised classifiers to localize the SZ against gold standard clinical localization. To further explore the prevalence of pathological features in an extended brain network outside of the clinically identified SZ, we also used unsupervised classification. Supervised SZ classification trained on individual features achieved accuracies of 0.6–.07 area under the receiver operating characteristic curve (AUC).* Combining all criticality and synchrony features further improved the AUC to 0.85. Unsupervised classification discovered an EpiNet-like cluster of brain regions, in which 51% of brain regions were outside of the SZ. Brain regions in the EpiNet-like cluster engaged in interareal hypersynchrony and locally exhibited high-amplitude bistability and excessive inhibition, which was strikingly similar to the high seizure risk regime revealed by our computational modeling. The finding that combining biomarkers improves SZ localization accuracy indicates that the novel mechanistic biomarkers for epileptogenicity employed here yield synergistic information. On the other hand, the discovery of SZ-like brain dynamics outside of the clinically defined SZ provides empirical evidence of an extended pathophysiological EpiNet.