ABSTRACT
Segmentation is the process of delineation of preferred region of interest and plays vital role in computer vision and medical field. This work focuses on the fully connected convolution neural network (CNN) for the segmentation of magnetic resonance (MR) brain images. The CNN is one of the popular deep learning architecture and has proved its efficiency in many real time applications. The volumetric 3D input is used in this work, and prior to segmentation, normalization was performed. The fully connected CNN architecture proposed in this work efficiently segments the gray matter, white matter and cerebrospinal fluid. Prior to segmentation, the preprocessing phase comprises normalization and the enhancement by adaptive histogram equalization. The 3D CNN architecture was tested on 20 3D MR data sets, and for validation, dice coefficient, Jaccard coefficient, false positive, false negative and true positive values. The average dice coefficient values of white matter (WM), gray matter (GM) and cerebrospinal fluid (CSF) extraction are 0.7638, 0.7658 and 0.7600. The performance metrics evaluation by metrics reveals the superiority of 3D CNN architecture in the segmentation of brain tissue components. The algorithms are developed in Matlab2015a and tested on real-time MR images of brain.
