3DBrainMiner, A tool to generate brain graphs from MR-images

The rise of brain imaging methods generates a considerable mass of morphological and functional data. However, their exploration and comparison over time for an individual (development and aging), between individuals (variability within the species), and even more so between different species, have been only shortly studied One strategy to facilitate this exploration is to first model this data with graphs. These brains can then be analyzed using classical graph theory methods (e.g., graph mining and graph-matching) or more recent approaches of artificial intelligence (deep neural networks on graphs, geometric deep learning, etc.).The most common way to model a brain as a graph consists in representing brain structures as nodes and anatomical and/or functional connections between structures as edges. This representation, therefore, requires a good knowledge of the species under consideration, which is rarely available across the entire animal kingdom. In this context, we propose to create a new tool that allows specialists to easily transform/model 3D brain MRI images into graphs. This tool lets also specialists defining and automatically computing other nodes and edges attributes to better represent brain morphology. This tool named 3dBrainMiner can extract geometric and signal information from MRI images and the associated segmentations. To generate representations in the form of graphs, 3dBrainMiner takes the MRI image, the corresponding segmentation, the segmentation label list, and the connectivity matrix, if available. If the connectivity matrix is not provided, 3dBrainMiner constructs edges in the graph by connecting two structures if they are sufficiently close to each other. Based on literature data, we know that the size of certain structures varies due to living conditions, pathologies, etc. Furthermore, observations of brains from different animal species show that structures can have different shapes (e.g. the caudate nucleus in sea lions etc.). In this context, we have chosen to introduce new descriptors without functional biases. For the nodes, 3dBrainMiner calculates shape descriptors (volume, surface, sphericity), the position of the center of gravity of each structure, and extracts information on gray-level intensities (mean, variance, radial profile). All this information is included in the graph represention with help of node attributes. Concerning edge attributes, the tool calculates distances (minimum, maximum, mean) and contact surfaces between connected nodes.Using this tool, we have constructed brain graphs for quails, lambs, and humans and have started analyzing the obtained graphs. In particular, for the quail model, we have initiated work on lineage classification using graph neural networks. 3DBrain miner integrate a computer user interface that allow to display at the same time the 3D MRI image and the resulting graph. 3DBrainminer source code is available in open access at URL : https://scm.univ-tours.fr/projetspublics/lifat/3dbrainminer