Brain Tumor Augmentation Using the U-NET Architecture

Abstract

Studies have found out that tumors in brain are one of the fiercest diseases which can ultimately lead to death. Gliomas are the most commonly found primary tumors that are very hard to predict and can be found anywhere in the brain. It is prime objective to differentiate the different tumor tissues such as enhancing tissues, edema, from healthy ones. To do this task, two types of segmentation techniques come into existent i.e. manual and automatic. The automation methods of brain tumor segmentation have gained ground over manual segmentation algorithms and further its estimation is very closer to clinical results. In this paper we propose a comprehensive U-NET architecture with modification in their layers for 2D slices segmentation as a major contribution to BRATS 2015 challenge.. Then we enlisted different datasets that are available publicly i.e. BRATS and DICOM. Further, we present a robust framework inspired from U-NET model with addition and modification of layers and image pre-processing methodology such as contrast enhancement for visible input and output details. In this way our approach achieves highest dice score 0.92 on the publicly available BRATS 2015 dataset and with better time constraint i.e. training time decreases to 80-90 minute instead of previously 2 to 3 days. We put our approach to the test on the benchmark brats 2015 dataset, and it outperformed the competition in terms of performance and Dice Score.