Learning from Enhanced Contextual Similarity in Brain Imaging Data for Classification of Schizophrenia

In certain severe mental diseases, like schizophrenia, structural alterations of the brain are detectable by magnetic resonance imaging (MRI). In this work, we try to automatically distinguish, by using anatomical features obtained from MRI images, schizophrenia patients from healthy controls. We do so by exploiting contextual similarity of imaging data, enhanced with a distance metric learning strategy (DML - by providing “must-be-in-the-same-class” and “must-not-be-in-the-same-class” pairs of subjects). To learn from contextual similarity of the subjects brain anatomy, we use a graph-based semi-supervised label propagation algorithm (graph transduction, GT) and compare it to standard supervised techniques (SVM and K-nearest neighbor, KNN). We performed out tests on a population of 20 schizophrenia patients and 20 healthy controls. DML+GT achieved a statistically significant advantage in classification performance (Accuracy: 0.74, Sensitivity: 0.79, Specificity: 0.69, Ck: 0.48). Enhanced contextual similarity improved performance of GT, SVM and KNN offering promising perspectives for MRI images analysis.