STUDY OF BRAIN REGION SEGMENTATION USING CONVOLUTIONAL NEURAL NETWORK

Abstract

Magnetic Resonance Imaging (MRI) is used in medical imaging for detection of tumours and visualize brain tissues.This is done manually by expert radiologist and this takes good amount of time. The traditional method of MRI evaluation of tumour depends greatly on qualitative features,like density of tumour,growth pattern etc. Brain Region Segmentation is important in neuroimaging application, for example, alignment of images, surface reconstruction etc.The previous methods depends upon the qualitative features and is very sensitive to errors. Noise and errors need to be reduced and efficiently delineated, very less work is done in automatic tumour detection using deep learning methods and there is lot of areas which can be explored. The deep learning method is very much different from the machine learning method. The machine learning method uses algorithms to input data, learn from given data, and make decision based on the experience or learning whereas the deep learning can learn and make decisions on its own. Deep learning has a capability of learning from data that is unstructured or unlabeled. In deep learning, the algorithms try to learn using method of feature extraction which is very different and makes the model fully automatic, here we don’t require any handcrafted feature. In traditional method we need to develop feature extractor for different problem, so we use deep learning which reduces effort of developing different feature extractor for different problem.

In this project we made a literature review on the existing methods used for segment brain tumour for exploring the most efficacious approach and decided to use Convolutional Neural Network (CNN).CNN learns features from the input image with the help of supervised learning or unsupervised learning. In one of method of 2-D patch extraction could achieve accuracy of 88% where the network architecture is inspired by VGG Network, high grade and low grade network differs in number of convolutional layer preceding a max-pooling layer. In other,they have used encoder-decoder type neural network and achieved accuracy of 87.2%. In a single forward pass, previously discussed patch based technique are slow as network predicts only centre pixel of patch. In the present study, we have used supervised learning to learn the features from the input images and found that Convolutional 2 Neural Network can achieve good accuracy.In CNN, the network in starting phase learns low level feature like lines or edges and then slowly learns the high level features. The present method achieved accuracy of 90-94% which is a good achievement in this field. MICCAI-BRATS challenge 2015 dataset is utilized in the present study. In present method, there are total of 245 MRI images which are further divided into 110 image for training the network and 145 images for testing the data.