King's College London

Research portal

Discriminative log-euclidean kernels for learning on brain networks

Research output: Chapter in Book/Report/Conference proceedingOther chapter contribution

Original languageEnglish
Title of host publicationConnectomics in NeuroImaging - 1st International Workshop, CNI 2017 Held in Conjunction with MICCAI 2017, Proceedings
PublisherSpringer Verlag
Pages25-34
Number of pages10
Volume10511 LNCS
ISBN (Print)9783319671581
DOIs
StateE-pub ahead of print - 2 Sep 2017
Event1st International Workshop on Connectomics in NeuroImaging, CNI 2017 held in conjunction with the 20th International Conference on Medical Image Computing and Computer Assisted Intervention, MICCAI 2017 - Quebec City, Canada
Duration: 14 Sep 201714 Sep 2017

Publication series

NameLecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume10511 LNCS
ISSN (Print)03029743
ISSN (Electronic)16113349

Conference

Conference1st International Workshop on Connectomics in NeuroImaging, CNI 2017 held in conjunction with the 20th International Conference on Medical Image Computing and Computer Assisted Intervention, MICCAI 2017
CountryCanada
CityQuebec City
Period14/09/201714/09/2017

King's Authors

Abstract

The increasing availability of functional Magnetic Resonance Imaging (fMRI) has led to a number of studies of brain networks with the aim of developing computer aided diagnosis of disease. Typically these are based on a statistical or machine learning method operating on connectivity networks, or features derived from them. This work presents a novel kernel method allowing classification tasks on connectivity networks represented as symmetric positive definite (SPD) matrices. It defines a kernel based on geodesic distances measured on the Riemannian manifold of SPD matrices, and automatically adjusts the eigenvalues of the matrices to improve accuracy. This is coupled with a Gaussian Process (GP) classifier, and used to discriminate healthy controls from Schizophrenia patients. The new kernel offers superior classification accuracy to previous kernels, and the adjusted eigenvalues allow discovery of clinically meaningful differences in connectivity between patients and controls.

View graph of relations

© 2018 King's College London | Strand | London WC2R 2LS | England | United Kingdom | Tel +44 (0)20 7836 5454