Machine Learning for Tract Segmentation in dMRI data

Diffusion MRI (dMRI) data allows to reconstruct the 3D pathways of axons within the white matter of the brain as a set of streamlines, called tractography. A streamline is a vectorial representation of thousands of neuronal axons expressing structural connectivity. An important task is to group the same functional streamlines into one tract segmentation. This work is extremely helpful for neuro surgery or diagnosing brain disease. However, the segmentation process is difficult and time consuming due to the large number of streamlines (about 3 × 10 5 in a normal brain) and the variability of the brain anatomy among different subjects. In our project, the goal is to design an effective method for tract segmentation task based on machine learning techniques and to develop an interactive tool to assist medical practitioners to perform this task more precisely, more easily, and faster. First, we propose a design of interactive segmentation process by presenting the user a clustered version of the tractography in which user selects some of the clusters to identify a superset of the streamlines of interest. This superset is then re-clustered at a finer scale and again the user is requested to select the relevant clusters. The process of re-clustering and manual selection is iterated until the remaining streamlines faithfully represent the desired anatomical structure of interest. To solve the computational issue of clustering a large number of streamlines under the strict time constraints requested by the interactive use, we present a solution which consists in embedding the streamlines into a Euclidean space (call dissimilarity representation), and then in adopting a state-of-the art scalable implementation of the k-means algorithm. The dissimilarity representation is defined by selecting a set of streamlines called prototypes and then mapping any new streamline to the vector of distances from prototypes. Second, an algorithm is proposed to find the correspondence/mapping between streamlines in tractographies among two different samples, without requiring any transformation as the traditional tractography registration usually does. In other words, we try to find a mapping between the tractographies. Mapping is very useful for studying tractography data across subjects. Last but not least, by exploring the mapping in the context of dissimilarity representation, we also propose the algorithmic solution to build the common vectorial representation of streamlines across subject. The core of the proposed solution combines two state-of-the-art elements: first using the recently proposed tractography mapping approach to align the prototypes across subjects; then applying the dissimilarity representation to build the common vectorial representation for streamline. Preliminary results of applying our methods in clinical use-cases show evidence that our proposed algorithm is greatly beneficial (in terms of time efficiency, easiness.etc.) for the study of white matter tractography in clinical applications.