A Region-Based Feature Extraction Method for Rs-fMRI of Depressive Disorder Classification

摘要

Depressive Disorder (DD) is the major disease among mood disorders, which has a high rate of misdiagnosis. Aiming at the fact that the feature dimension of brain image is much higher than the sample size. What it needs to be solved is that the locally valid details might be ignored when extracting global features. Thus, in this paper, a machine learning classification method based on regional feature extraction is proposed. First, the brain image is equally divided into 27 regions, and then patches are extracted from each local region and clustered into clusters. A Filter-based feature selection is performed on the patch and fused to the regional features. Then, an Embedded-based feature selection is used to extract the weights of regions which are connected as global features. Finally, the SVM is adopted for classification. The method is tested on resting-state functional Magnetic Resonance Imaging (rs-fMRI) of 46 DD patients and 46 Healthy Control (HC) subjects captured by a psychiatric hospital and compared with existed methods. Our proposed method achieved 95.65% accuracy and 91.30% recall for classification of DD vs. HC, whose performance is higher than those of available methods, demonstrating the effectiveness of classification.CCS Concepts•Signal Processing $ightarrow$Bioimaging and Signal Processing•Applications of Signal Processing $ightarrow$ Machine Learning for Signal Processing.