Brain tumor segmentation and multiview multiscale‐based radiomic model for patient's overall survival prediction

Abstract

AbstractA brain tumor is the most common primary brain malignancy. Delaying in brain tumor diagnosis is a primary cause of death in affected individuals. Therefore, early diagnosis of a brain tumor is essential for treatment planning and prognosis. In this study, the multilevel dilated convolutional neural network (MLDCNN) model is used for brain tumor segmentation. MLDCNN model is implemented independently for five MLDC blocks with a different combination of dilation rates to analyze their impact on brain tumor segmentation. For each segmentation model, overall survival time prediction is performed independently. An automated system is proposed for the overall survival time prediction of patients suffering from a brain tumor. First, shape and multiscale texture‐based features are extracted from LoG filtered and wavelet decomposed images of the magnetic resonance imaging scans. The proposed model utilizes 3D information by extracting radiomic features from axial, coronal, and sagittal views. These features are reduced using an extra tree classifier to avoid overfitting. Random forest algorithms are applied on selected feature sets to predict overall survival time in days. Extensive experimentation is performed for the segmentation and survival time prediction on the publicly available BraTS2019 and BraTS 2020 datasets. Results demonstrate that the proposed approach achieved the least mean squared error value in the survival time prediction task.