SegNet-based Corpus Callosum segmentation for brain Magnetic Resonance Images (MRI)

Abstract

Corpus callosum is the most significant human brain structures. The majority of neurological disorder directly or indirectly reflect on Corpus Callosum morphological characteristics. The mid-sagittal view of the Tl weighted brain MRI completely portray corpus callosum anatomical structure. The segmentation of corpus callosum from brain MRI is very challenging task due to low contrast in surrounding organ and tissues. We propose a novel Corpus Callosum segmentation method using semantic pixel-wise segmentation termed as SegNet, a practical deep convolutional neural network architecture. The applied architecture comprises of two networks namely encoder and decoder with pixel-specific classification layer. The proposed model’s encoder network comprises of series of convolution, batch normalization and max-pool layers. The function of decoder network is to map the feature maps of the low-resolution encoder to the full input resolution featuremaps for the classification of pixels. The segmentation output can be used for better extraction of features and classification of diseases in medical diagnosis.