Brain Tumor Segmentation Evaluation and Overall Survival Prediction on the BRATS 2020 Dataset

Abstract

Overall survival (OS) assessment in glioma patients is a crucial component of MRI-based medical image analysis, as OS estimation directly influences clinical decision-making and treatment planning. One of the central challenges in developing image-based predictive models lies in the dependency on accurate tumor segmentation. This study aims to construct an MRI-based OS prediction model using the Brain Tumor Segmentation 2020 (BraTS 2020) dataset by incorporating two types of masked images: ground truth masks and automatically generated predicted masks derived from a 3D U-Net segmentation model. OS classification was grouped into three categories (< 10 months, 10–15 months, and > 15 months). The predictive model achieved an accuracy of 0.9792 when using ground truth masks and 0.9583 when using predicted masks. These findings suggest that a fully automated deep-learning–based segmentation pipeline can approximate the performance of manual segmentation and holds strong potential for large-scale clinical applications where manual annotation is impractical.