Brain Tumor Detection and Segmentation in MR Images Using Deep Learning

Gliomas are the most infiltrative and life-threatening brain tumors with exceptionally quick development. Gliomas segmentation using computer-aided diagnosis is a challenging task, due to irregular shape and diffused boundaries of tumor with the surrounding area. Magnetic resonance imaging (MRI) is the most widely used method for imaging structures of interest in human brain. In this study, a deep learning-based method that uses different modalities of MRI is presented for the segmentation of brain tumor. The proposed hybrid convolutional neural network architecture uses patch-based approach and takes both local and contextual information into account, while predicting output label. The proposed network deals with over-fitting problem by utilizing dropout regularizer alongside batch normalization, whereas data imbalance problem is dealt with by using two-phase training procedure. The proposed method contains a preprocessing step, in which images are normalized and bias field corrected, a feed-forward pass through a CNN and a post-processing step, which is used to remove small false positives around the skull portion. The proposed method is validated on BRATS 2013 dataset, where it achieves scores of 0.86, 0.86 and 0.91 in terms of dice score, sensitivity and specificity for whole tumor region, improving results compared to the state-of-the-art techniques.