CNN and Ensemble Learning for Brain Tumor Segmentation and Survival Prediction

Abstract

Brain tumor is the spread of abnormal cells in the brain. Out of several kinds of brain tumor gliomas is the most dangerous with low survival rate and difficult to detect manually due to irregular form and confusing boundaries. Magnetic Resonance Imaging is the most widely used imaging modality that allows radiologist to look inside brain by utilizing radio waves and magnet but the manual identification of tumor region is tedious task. Therefore, a reliable and automatic segmentation and prediction is necessary for segmentation of brain tumor and prediction. However due to complexity and unavailability of resources to train deep learning algorithms, it is complex to identify the tumorous and non-tumorous region. So, in this paper, a reliable and efficient variant of UNET i.e., VGG19-UNET for segmentation of brain tumor and ensemble learning model for survival prediction is proposed. Specifically, an encoder part of the UNET is a pretrained VGG19 followed by the adjacent decoder part. Meanwhile, the ensemble voting classifier of Naïve Bayes and Random Forest was trained for survival prediction. The datasets we are using for segmentation is BRATS’20 which comprises of four different MRI modalities and one target mask file. Whereas, the datasets of survival prediction is also BTARS’20 which is comma separated file containing different features. Above mentioned algorithm resulted in dice coefficient score of 0.81, 0.86 and 0.88 for enhancing, core and whole tumor whereas the accuracy of overall survival is 62.7%.