Abstract:Purpose Brain tumor is one of the most dangerous and life-threatening disease. In order to decide the type of tumor, devising a treatment plan and estimating the overall survival time of the patient, accurate segmentation of tumor region from images is extremely important. The process of manual segmentation is very time-consuming and prone to errors; therefore, this paper aims to provide a deep learning based method, that automatically segment the tumor region from MR images. Design/methodology/approach In this paper, the authors propose a deep neural network for automatic brain tumor (Glioma) segmentation. Intensity normalization and data augmentation have been incorporated as pre-processing steps for the images. The proposed model is trained on multichannel magnetic resonance imaging (MRI) images. The model outputs high-resolution segmentations of brain tumor regions in the input images. Findings The proposed model is evaluated on benchmark BRATS 2013 dataset. To evaluate the performance, the authors have used Dice score, sensitivity and positive predictive value (PPV). The superior performance of the proposed model is validated by training very popular UNet model in the similar conditions. The results indicate that proposed model has obtained promising results and is effective for segmentation of Glioma regions in MRI at a clinical level. Practical implications The model can be used by doctors to identify the exact location of the tumorous region. Originality/value The proposed model is an improvement to the UNet model. The model has fewer layers and a smaller number of parameters in comparison to the UNet model. This helps the network to train over databases with fewer images and gives superior results. Moreover, the information of bottleneck feature learned by the network has been fused with skip connection path to enrich the feature map.

Deep learning-based approach for segmentation of glioma sub-regions in MRI