Unsupervised Brain Segmentation System Using K-Means and Neural Network

Abstract

AbstractVoxel-based morphometry is an analysis technique used to quantify matter volume in the human brain from magnetic resonance imaging studies. Atrophies and morphologic changes can be signs of neuronal depletion that can lead to various degenerative diseases. Two of the most worldwide used tools for brain volume assessment and segmentation in white matter, grey matter, and cerebrospinal fluid are the Functional Magnetic Resonance Imaging of the Brain (FMRIB) Software Library (FSL) and Statistical Parameter Mapping (SPM). However, the main issue of these tools is related to the choice of the best parameter setting.In this paper, a novel and unsupervised segmentation system, without any user parameter setting, that uses both the k-means clustering algorithm and artificial neural network is proposed. Performance of this system was evaluated on the Internet Brain Segmentation Repository (IBSR) dataset (v.2.0) and results were compared with FSL and SPM results.The dice similarity score was calculated to compare the segmentations obtained with FSL, SPM, and the proposed system with the reference segmentations. The proposed system resulted fast and reliable, and reduced any error (the dice similarity score was greater than 83%, 86%, and 65% for white matter, grey matter, and cerebrospinal fluid, respectively) demonstrating an improvement in the discrimination among white matter, grey matter, and cerebrospinal fluid.KeywordsVoxel-Based MorphometryK-meansArtificial neural networks