Learning from Diffusion-Weighted Magnetic Resonance Images using graph kernels

Sylvain Takerkart (1) , Gottfried Berton (1) , Nicole Malfait (1) , François-Xavier Dupé (2)

1 INT - Institut de Neurosciences de la Timone
2 QARMA - éQuipe AppRentissage et MultimediA [Marseille]

Diffusion-weighted magnetic resonance imaging (DWI) is a scanning procedure that allows infering the anatomical connectivity of the brain non invasively. DWI can be used to segment the brain into a set of relevant sub-regions, yielding what is called a parcellation in the neuroimaging literature. In this paper, we introduce a generic framework that allows building predictive models using parcellations obtained on a single individual. It consists in constructing attributed region adjacency graphs to represent the parcellations and using suitable graph kernels to exploit the versatility of kernel methods. We demonstrate the relevance of this framework on real data, by showing that we can predict the age range of an individual from the connectivity structure of its corpus callosum, the main hub of connections between the left and right hemispheres of the brain. Furthermore, we study the behavior of different graph kernels for this task. This work opens new opportunities to identify DWI-based biomarkers of neurodegenerative and psychiatric diseases.

Mots clés: region-adjacency graphs; graph kernel; magnetic resonance imaging; brain connectivity; brain parcellation

Domaines: Neurosciences [q-bio.NC]; Imagerie médicale; Traitement du signal et de l'image [eess.SP]; Apprentissage [cs.LG]

Sylvain Takerkart, Gottfried Berton, Nicole Malfait, François-Xavier Dupé. Learning from Diffusion-Weighted Magnetic Resonance Images using graph kernels. 11th IAP Workshop on Graph-based Representation, GbR2017, May 2017, Anacapri, Italy. ⟨hal-01484660⟩