Multiphysics image-based AnalysIs for premature brAin development understanding – MAIA

The MAIA project aims to develop innovative and robust solutions for the analysis of signals/images of the premature brain, with the aim of designing effective tools to improve our understanding of neurological development.
The first challenge is related to the size and quantity of data. This involves developing automatic approaches that can take advantage of IT resources and rely on appropriate algorithmic frameworks.
In response to this challenge, the proposed approach is to significantly improve the processing of morphological data such as image enhancement, super-resolution and brain segmentation to provide robust morphometric analysis tools, leading to improved knowledge of early in vivo brain growth.
The second challenge concerns the signal and images coming from different physical phenomena (magnetic, electrical, optical). Reliable brain function measurements combined with accurate localization are needed for neonatal brain mapping. Our approach focuses on optimizing the acquisition and analysis of functional data specific to the newborn. In particular, we will focus on the acquisition of ASL, which remains unexplored for newborns, as well as the rapid and reliable resolution of inverse problems for EEG and NIRS data. A major improvement in the analysis of neonatal brain function is needed to improve our understanding of the impact of early brain damage on children's development.
Finally, brain structures are subject to various changes, from macroscopic to microscopic scales. The approach followed will be to develop patient-specific analysis techniques to take into account the anatomical and functional variability of each patient.

A first method has been proposed to reconstruct a high-resolution image from a single low-resolution image («single image super resolution«). The approach is based on a deep convolutional neural network («Deep Learning«). The results, published at the IEEE ISBI' 2017 international conference, showed the relevance of the proposed approach for adult MRI images, and it is planned to continue this work for the MAIA neonatal images.

With regard to segmentation of brain MRI data, several approaches have been developed: 1) morphologically-based 3D superpixel pre-segmentation, 2) interactive segmentation based on partition binary trees, and 3) a linear multi-atla patches approach for cortex segmentation.

In the MAIA project, we also have an iterative approach to constructing 4D atlases and a method for improving the robustness of registration algorithms when mapping brains of different sizes has been implemented.

Un des points clefs du projet concerne la mise en plateforme dédiée pour faciliter l’analyse des données. Nous procéderons ainsi à l’implantation d’un pont entre Shanoir et la plateforme de données de recherche afin de permettre de garder les informations cliniques sur site et de donner l'accès pour le traitement d’image.

The perspectives concern the aspects of robustness with regard to the movement of premature babies during acquisitions, but also the pooling of developments carried out to set up an MRI data processing pipeline for the analysis of functional data.

Regarding the reconstruction of MRI images, we are moving towards the development of a technique for the correction of slice-to-slice motion without using an intermediate reference image (which is the main defect of current reconstruction methods). The development of the approaches studied for MRI data segmentation will be continued in order to provide a robust toolbox for the research community.

One of the key points of the project concerns the implementation of a dedicated platform to facilitate data analysis. We will thus proceed with the implementation of a bridge between Shanoir and the research data platform in order to keep clinical information on site and provide access for image processing.

C.-H. Pham, A. Ducournau, R. Fablet, F. Rousseau . Brain MRI super-resolution using deep 3D convolutional networks. IEEE Int. Symposium on Biomedical Imaging. April 2017

Each year in France, 55 000 children are born prematurely, i.e., before the 37th week of gestation. In this context, extreme prematurity (less than 32 weeks) covers 10 000 newborns a year. Long-term studies of the outcome of prematurely born infants have clearly documented that the majority of such infants may have significant motor, cognitive, and behavioral deficits.

However, there is a limited understanding of the nature of the cerebral abnormality underlying these adverse neurologic outcomes. The French EPIPAGE study shows that nearly a third of these former preterm infants still require specific cares at 5 years-old. Researchers conclude that “prevention of the learning disabilities associated with cognitive deficiencies [...] is an important goal for modern perinatal care for children who are born very preterm and for their families”.

A first answer to this challenge was proposed by 3D morphological imaging, namely Magnetic Resonance Imaging (MRI). It led to recent breakthroughs, related to the potential correlations between specific tissue lesions, or volumetric variations of given cerebral structures, on the one hand, and some disabilities, mainly motor, on the other hand. Nevertheless, the understanding of the intrinsic mechanisms resulting in cognitive disabilities remains limited without having access to functional information.

In this context, the emergence of new modalities of 3D functional MRI (e.g., Arterial Spin Labeling, ASL), or optical imaging technologies (e.g., Near InfraRed Spectroscopy, NIRS), bring new perspectives for extracting cognitive information, via metabolic activity measures. Other classical technics devoted to cerebral signal measurement, such as ElectroEncephaloGraphy (EEG), provide cognitive information at the cortical level. Each of these various non-invasive imaging technologies brings substantial and specific information for the understanding of newborn brain development.

However, the induced data of preterm brain are voluminous, noisy and highly heterogeneous, in terms of nature (signal, image), dimensions (1D, 2D, 3D, time) and semantics (morphological, physiological, functional). As a consequence, establishing correspondences between these data and cross-analysing their underlying information remains so far a challenging task.

In order to tackle these challenges, this project aims at developing innovative approaches for multi-image / multi-signal analysis, in order to improve neurodevelopment understanding methods.

To reach such goal – that requires to handle fundamental, methodological and technological issues – both pluri- / interdisciplinary and mixed academic / industrial interactions are mandatory.

From a fundamental point of view, mathematics and computer science have to be considered in association with imaging physics and medicine, to deal with open issues of signal and image analysis from heterogeneous data (image, signal), considered in the multiphysics contexts related to data acquisition (magnetic, optic, electric signals) and biophysics modeling of the newborn brain. A sustained synergy between all these scientific domains is then necessary.

Finally, the sine qua non condition to reach a better understanding of the coupled morphological-cognitive development of premature newborns, is the development of effective software tools, and their distribution to the whole medical community. The very target of this project will be the design of such software tools for medical image / signal analysis, actually operational in clinical routine, and freely available. Academic researchers and industrial partners will work in close collaboration to reach that ambitious goal.