A STUDY ON DEEP CONVOLUTIONAL NEURAL NETWORKS FOR BRAIN TURNOR MRI SEGMENTATION

Abstract

With the advent of new technologies in the field of medicine, there is rising awareness of biomechanisms, and we are better able to treat ailments than we could before. One such field is computer vision, which, over the years has helped us in inderstanding our body and its mechanisms in a better way. Among various forms of cancer, brain tumors are the most aggressive as they lead to the highest percentage of fatalities. It is also the second most common type of cancer in children below 15 years of age. Thus, treatment planning is a key stage to improve the quality of life of oncological patients. Magnetic resonance imaging (MRI) is a widely used imaging technique to assess these tumors, but the large amount of data produced by MRI prevents manual segmentation in a reasonable time, limiting the use of precise quantitative measurements in the clinical practice. So, automatic and reliable segmentation methods are required; however, the large spatial and structural variability among brain tumors make automatic segmentation a challenging problem. In this thesis, we first build upon the state-of-the-art model, which used a deep convolutional neural network with small convolutional kernels of size 3×3, to segment the brain tumor regions. The use of small kernels allows designing a deeper architecture, besides having a positive effect against overfitting, given the fewer number of weights in the network. This architecture used small image patches for pixelwise classification. Our work builds upon this as the baseline network, and we perform several experiments aimed at CNN architectures to develop insights. We then try to understand qualitatively how the features were learned in this network, by visualizing the filter responses. Our work then progressed towards experimenting with similar architectures, and observing the effect of pooling operations on this network. We also studied an alternate architecture which used convolution layer instead of pooling. We also study the effects of global vs local features in image, by proposing a fully convolutional network which gives dense classification of pixels, and is faster than pixelwise patch-based classification.