Deep Learning based Model for Brain Tumor Detection and Segmentation using BRATS Dataset

This paper presents a method that can automatically perform brain tumor segmentation using Deep Neural Networks (DNNs). Glioblastomas (both low and high grade) pictured in MR images are used to tailor the proposed system. These tumors are varied in shape, size and contrast and can be found anywhere in the brain. Our exploration of an extremely efficient machine learning solution run on a flexible and high capacity DNN is inspired by this nature of the tumors. We have worked on various choices of models indoor to promise a competitive performance. The different architectures that we explored were based on Convolutional Neural Networks (CNN). These architectures are also called DNN and their key feature is their adaptability to image data. Our novel CNN architecture is different from the traditionally used ones in Computer Vision for the following reasons. Both local features and global features are simultaneously exploited. A convolutional implementation of a layer which is fully connected is used as a final layer in our networks. This speeds up the process by a 40 fold. Difficulties which relate to the imbalance of tumor labels are handled by introducing a 2-phase training procedure. In the end, a cascade architecture is explored. Here for a subsequent CNN, the output of a basic CNN is used as an additional source of information.