Overall Survial Prediction from Brain MRI in Glioblastoma

Abstract

Tumor segmentation using radiological images, particularly in the brain region, is a challenging task due to the heterogeneous nature of the tissue representing the brain lesions. In computerized diagnostic systems, this method is an essential step in isolating tumor regions for visualization and examination. Recently, deep learning (DL) technology has resulted in major breakthroughs in computer vision and artificial intelligence. This has impacted clinical tasks such as brain tumor segmentation, where deep learning allows learning hierarchical and distinctive characteristics from radiographs. A major paradigm shift has evolved, when compared with traditional machine learning approaches, where pathological and healthy tissue can be differentiated without relying on extracting features that required a significant expertise. Herein, with our proposed method, we aim to help radiologists in assisting them to detect tumor regions and further predict the overall patient survival rate using magnetic resonance images effectively. Towards this, we use U-Net based architecture to perform the segmentation. We achieve acceptable segmentation accuracy, 82% and 75% on training and validation datasets, respectively. We further used our segmentation results for survival predication task. We computed and selected 16 most significant 3D and 2D radiomic features from the segmented regions. By combining age with radiomics features, we trained Convolutional Neural Network (CNN) model and five different machine learning (ML) models and achieved 65.57% and 63% accuracy in survival rate prediction when using CNN and support vector machine (SVM) classification model.