Section: New Results

Research axis 1: Medical Image Computing in Neuroimaging

Extraction and exploitation of complex imaging biomarkers involve an imaging processing workflow that can be quite complex. This goes from image physics and image acquisition, image processing for quality control and enhancement, image analysis for features extraction and image fusion up to the final application which intends to demonstrate the capability of the image processing workflow to issue sensitive and specific markers of a given pathology. In this context, our objectives in the recent period were directed toward following major methodological topics:

Diffusion imaging

Free water estimation using single-shell diffusion-weighted images

Participant : Emmanuel Caruyer.

Free-water estimation requires the fitting of a bi-compartment model, which is an ill-posed problem when using only single-shell data. Its solution requires optimization, which relies on an initialization step. We propose a novel initialization approach, called "Freewater EstimatoR using iNtErpolated iniTialization" (FERNET), which improves the estimation of free water in edematous and infiltrated peritumoral regions, using single-shell diffusion MRI data. The method has been extensively investigated on simulated data and healthy and brain tumor datasets, demonstrating its applicability on clinically acquired data. Additionally, it has been applied to data from brain tumor patients to demonstrate the improvement in tractography in the peritumoral region [57].

Multi-dimensional diffusion MRI sampling scheme: B-tensor design and accurate signal reconstruction

Participant : Emmanuel Caruyer.

B-tensor encoding enables the separation of isotropic and anisotropic tensors. However, little consideration has been given as to how to design a B-tensor encoding sampling scheme. In this work, we propose the first 4D basis for representing the diffusion signal acquired with B-tensor encoding. We study the properties of the diffusion signal in this basis to give recommendations for optimally sampling the space of axisymmetric b-tensors. We show, using simulations, that the proposed sampling scheme enables accurate reconstruction of the diffusion signal by expansion in this basis using a clinically feasible number of samples [24].

This work was done in collaboration with A. Bates, Australian National University and Al. Daducci, University of Verona.

Optimal selection of diffusion-weighting gradient waveforms using compressed sensing and dictionary learning

Participants : Raphaël Truffet, Emmanuel Caruyer, Christian Barillot.

Acquisition sequences in diffusion MRI rely on the use of time-dependent magnetic field gradients. Each gradient waveform encodes a diffusion-weighted measure; a large number of such measurements are necessary for the in vivo reconstruction of microstructure parameters. We propose here a method to select only a subset of the measurements, while being able to predict the unseen data using compressed sensing. We learn a dictionary using a training dataset generated with Monte-Carlo simulations; we then compare two different heuristics to select the measures to use for the prediction. We found that an undersampling strategy limiting the redundancy of the measures allows for a more accurate reconstruction when compared with random undersampling with similar sampling rate [49].

Geometric evaluation of distortion correction methods in diffusion MRI of the spinal cord

Participants : Haykel Snoussi, Emmanuel Caruyer, Olivier Commowick, Benoit Combès, Élise Bannier, Christian Barillot.

Acquiring and processing Diffusion MRI in spinal cord present inherent challenges. Differences in magnetic susceptibility between soft tissues, air and bones make the magnetic field non uniform in spinal cord. In this context, various procedures were proposed for correcting inhomogeneity-induced distortions; in this work, we propose novel geometric statistics to measure the alignment of the reconstructed diffusion model with the apparent centerline of the spine. In parallel of the correlation with an anatomical T2-weighted image, we show the utility of these statistics to study and evaluate the impact of distortion correction by comparing three correction methods using a pair of images acquired with reversed gradient polarity [48].

This work was done in collaboration with Anne Kerbrat, Neuropoly Montréal and Julien Cohen-Adad from NeuroPoly Lab, Institute of Biomedical Engineering, Polytechnique Montreal, Montreal, QC, Canada.

Arterial Spin Labeling

Acquisition duration in resting-state arterial spin labeling. How long is enough?

Participants : Corentin Vallée, Pierre Maurel, Isabelle Corouge, Christian Barillot.

Resting-state Arterial Spin Labeling (rs-ASL) is a rather confidential method compared to resting-state BOLD but it drives great prospects with respect to potential clinical applications. By enabling the study of cerebral blood flow maps, rs-ASL can lead to significant clinical subject-scaled applications as CBF is a biomarker in neuropathology. An important parameter to consider in functional imaging is the acquisition duration. Despite directly impacting practicability and functional networks representation, there is no standard for rs-ASL. Our work here focuses on strengthening the confidence in ASL as a rs-fMRI method, and on studying the influence of the acquisition duration. To this end, we acquired a long rs-ASL sequence and assessed the quality of typical functional brain networks quality over time compared to gold-standard networks. Our results show that after 14min of duration acquisition, functional networks representation can be considered as stable [58], [50].

Patch-based super-resolution of arterial spin labeling magnetic resonance images

Participants : Cédric Meurée, Pierre Maurel, Jean-Christophe Ferré, Christian Barillot.

Arterial spin labeling is a magnetic resonance perfusion imaging technique that, while providing results comparable to methods currently considered as more standard concerning the quantification of the cerebral blood flow, is subject to limitations related to its low signal-to-noise ratio and low resolution. In this work, we investigated the relevance of using a non-local patch-based super-resolution method driven by a high-resolution structural image to increase the level of details in arterial spin labeling images. This method was evaluated by comparison with other image resolution increasing techniques on a simulated dataset, on images of healthy subjects and on images of subjects diagnosed with brain tumors, who had a dynamic susceptibility contrast acquisition. The influence of an increase of ASL images resolution on partial volume effects was also investigated in this work [16].

The development of this super-resolution algorithm in the context of the PhD of Cédric Meurée founded by Siemens Healthineers conducted to a stay of one month of the PhD candidate in Erlangen, during summer 2018. This immersion into the neuro-development team allowed him to integrate the proposed solution with tools in use within this team. Part of the work also consisted in reducing the computation, a factor of 5 being achieved at the end of these four weeks.

Atlases

Unbiased longitudinal brain atlas creation using robust linear registration and log-Euclidean framework for diffeomorphisms

Participants : Antoine Legouhy, Olivier Commowick, Christian Barillot.

We have defined a new method to create a diffeomorphic longitudinal (4D) atlas composed of a set of 3D atlases each representing an average model at a given age. This is achieved by generalizing atlasing methods to produce atlases unbiased with respect to the initial reference up to a rigid transformation and ensuring diffeomorphic deformations thanks to the Baker-Campbell-Hausdorff formula and the log-Euclidean framework for diffeomorphisms. Subjects are additionally weighted using an asymmetric function to closely match specified target ages. Creating a longitudinal atlas also implies dealing with subjects with large brain differences that can lead to registration errors. This is overcome by a robust rigid registration based on polar decomposition. We illustrated these techniques for the creation of a 4D pediatric atlas, showing their ability to create a temporally consistent atlas [22].

This work was done in collaboration with François Rousseau, IMT Atlantique, LaTIM U1101 INSERM, Brest, France, under the ANR MAIA project.

Online atlasing using an iterative centroid

Participants : Antoine Legouhy, Olivier Commowick, Christian Barillot.

Online atlasing, i.e. incrementing an atlas with new images as they are acquired, is key when performing studies on databases very large or still being gathered. We proposed to this end a new diffeomorphic online atlasing method without having to perform again the atlasing process from scratch. New subjects are integrated following an iterative procedure gradually shifting the centroid of the images to its final position, making it computationally cheap to update regularly an atlas as new images are acquired (only needing a number of registrations equal to the number of new subjects). We evaluated this iterative centroid approach through the analysis of the sharpness and variance of the resulting atlases, and the transformations of images, comparing their deviations from a conventional method using Guimond's method. We demonstrated that the transformations divergence between the two approaches is small and stable, and that both atlases reach equivalent levels of image quality [42].

This work was done in collaboration with François Rousseau, IMT Atlantique, LaTIM U1101 INSERM, Brest, France, under the ANR MAIA project.

Neurofeedback

Learning bi-modal EEG-fMRI neurofeedback to improve neurofeedback in EEG only

Participants : Claire Cury, Pierre Maurel, Giulia Lioi, Christian Barillot.

In neurofeedback (NF), a new kind of data are available: electroencephalography (EEG) and functional magnetic resonance imaging (fMRI) acquired simultaneously during bi-modal EEG-fMRI neurofeedback. These two complementary techniques have only recently been integrated in the context of NF for brain rehabilitation protocols. Bi-modal NF (NF-EEG-fMRI) combines information coming from two modalities sensitive to different aspect of brain activity, therefore providing a higher NF quality. However, the use of the MRI scanner is cumbersome and exhausting for patients. We presented, a novel methodological development, able to reduce the use of fMRI while providing to subjects NF-EEG sessions of quality comparable to the bi-modal NF sessions. We proposed an original alternative to the ill-posed problem of source reconstruction. We designed a non-linear model considering different frequency bands, electrodes and temporal delays, with a structured sparse regularisation. Results show that our model is able to significantly improve the quality of NF sessions over what EEG could provide alone. We tested our method on 17 subjects that performed three NF-EEG-fMRI sessions each [30].

Can we learn from coupling EEG-fMRI to enhance neuro-feedback in EEG only?

Participants : Claire Cury, Pierre Maurel, Christian Barillot.

Neurofeedback (NF) measures brain activation during a task, and gives back to the subject a score reflecting his/her performance that he/she tries to improve. Among noninvasive functional brain imaging modalities, the most used in NF, are electro-encephalography (EEG) and the functional magnetic resonance imaging (fMRI). EEG measures the electrical activity of the brain through channels located on the scalp, with an excellent temporal resolution (milliseconds), but has a limited spatial resolution due to the well-known ill-posed inverse problem of source reconstruction. Also NF-EEG (NF session with NF scores extracted from EEG) is not easy to control since it comes from mixtures of propagating electric potential fluctuations. Blood oxygenation level dependent (BOLD) fMRI measures neuro-vascular activity, easier to control, with an excellent spatial resolution, making NF-fMRI (NF session with NF scores extracted from BOLD-fMRI) an adequate modality for NF. However its temporal resolution is only of a few seconds, and it is a costly, exhausting for subjects and time consuming modality. Since those modalities are complementary, their combined acquisition is actively investigated, as well as the methodology to extract information from fMRI with EEG which is the easiest modality to use [Abreu et al. 2018]. Our challenge is to learn EEG activation patterns from NF-fMRI scores extracted during a NF session using coupled EEG-fMRI data (NF-EEG-fMRI) to improve NF scores when using EEG only [29].

Deep learning

Unsupervised domain adaptation with optimal transport in multi-site segmentation of multiple sclerosis lesions from MRI data

Participants : Antoine Ackaouy, Olivier Commowick, Christian Barillot, Francesca Galassi.

Automatic segmentation of Multiple Sclerosis (MS) lesions from Magnetic Resonance Imaging (MRI) images is essential for clinical assessment and treatment planning of MS. Recent years have seen an increasing use of Convolutional Neural Networks (CNNs) for this task. Although these methods provide accurate segmentation, their applicability in clinical settings remains limited due to a reproducibility issue across different image domains. MS images can have highly variable characteristics across patients, MRI scanners and imaging protocols. Retraining a supervised model with data from each new domain is not a feasible solution because it requires manual annotation from expert radiologists. In this work, we explored an unsupervised solution to the problem of domain shift. We presented a framework, Seg-JDOT, which adapts a deep model so that samples from a source domain and samples from a target domain sharing similar representations will be similarly segmented. We evaluated the framework on a multi-site dataset, MICCAI 2016, and showed that the adaptation towards a target site can bring remarkable improvements in a model performance over standard training [54].

This work was done in collaboration with Nicolas Courty, Obelix team, IRISA laboratory from University of Bretagne Sud.

Deep learning for multi-site MS lesions segmentation: two-step intensity standardization and generalized loss function.

Participants : Francesca Galassi, Olivier Commowick, Christian Barillot.

We presented an improved CNN framework for the segmentation of Multiple Sclerosis (MS) lesions from multi-modal MRI. It uses a two-step intensity normalization and a cascaded network with cost sensitive learning. Performance was assessed on a public multi-site data-set [35].