Abstract:
A Brain tumor is an abnormal cell growth in the brain tissues, these tumors are difficult to
treat and severely affect the patient’s cognitive ability. Out of all brain tumors, gliomas are the
deadliest with least survival rate. Glioma is one of the most common primary cancerous brain
tumors. They are the most aggressive kind of cancer therefore; a better treatment and planning is
crucial for the patient’s overall survival. Before starting a treatment, it is essential to correctly
differentiate healthy and cancerous tissues of the patient’s brain. Both manual and automatic
segmentation methods are utilized to segment the glioma brain tumors. With the advent of new
approaches, automatic segmentation processes are becoming more effective and clinically
accepted. The focus of automatic brain tumor segmentation task is to separate tumor tissue i.e.
edema, tumor core from the healthy tissues i.e. white cells, Cerebrospinal Fluid and gray matter.
We have developed a novel automatic segmentation framework consisting of ResNet architecture
which is based on Deep Convolutional Neural Network (DCNN). Deep Convolutional Neural
Networks (DCNN) consists of various layers i.e. convolution, pooling, activation, normalization
and fully connected layers. The extra number of layers helps in learning more abstract features of
the input. We utilized two-phase training in order to tackle the class imbalance problem in the
dataset. Furthermore, we studied various loss function optimizer to fine tune our results. We tested
our framework on a benchmark brats 2015 dataset where it achieved state-of-the-art performance
and achieved better results on a Dice Score.
