Application of deep learning in brain tumor segmentation

Abstract

Glioma is one of the most common brain tumor worldwide. Nowaday, glioma identification and segmentation is essential when making clinical decisions. Obviously, manual segmentation is not only time consuming but also subjective, in addition to this task, it is rather difficult to solve the solution of the automated segmentation methods. U-net architecture is one of the popular deep learning models for segmentation of biomedical images. This study implements a 2 dimensional convolutional neural network based on a U-net architecture. The proposed model is trained only with 2D images (slices) from 3D MRI dataset to segment the different tumor regions. Futhermore, in the preprocessing step some methods are applied to deal with imbalanced data for three brain tumor regions, and then data augmentation is utilized to prevent overfitting when training with a small dataset. The evaluated results based on the Multimodal Brain Tumor Segmentation Challenge (BraTS) 2018 dataset get the high accuracy for Dice scores of whole tumor, tumor core and enhancing tumor being 0.88, 0.81 and 0.76, respectively. The performance of the proposed model show that the brain tumors can be recognized with high accuracy compared to the similar studies. The proposed study has provided an analytical procedure for glioma brain tumors, from preprocessing to the training model, and then assessing factors that may affect the performance of the achived results.