ABSTRACT
In this paper, a study is reported on the popular BraTS dataset for segmentation of brain tumor. The BraTS 2019 dataset is used that comprises four MR modalities along with the ground-truth for 259 high grade glioma (HGG) and 76 low grade glioma (LGG) patient data. We have employed U-Net architecture based 2D convolutional neural network (CNN) for each of the orthogonal planes (sagittal, coronal and axial) and fused their predictions. The objective function is aimed to minimize Dice loss between the binary prediction and its actual labels. Samples having tumor information are considered for each patient data to avoid training on non-informative data. The models are trained on 222 HGG data and tested on 37 HGG data using performance metrics such as sensitivity, specificity, accuracy and Dice score. Test-time augmentation is also performed to improve the segmentation performance. 7-fold cross validation is conducted to analyze the performance on different sets of training and testing data.
