Latent Radiomic-Driven U-Net for Brain Tumor Segmentation, Survival Forecasting, and 3D Visualization

Brain tumor segmentation from Magnetic Resonance Imaging (MRI) is a challenging task due to the vast amount of data, physician shortage, high population, and time-consuming nature of the process. To address these challenges, we propose a method that utilizes a latent radiomic-driven U-Net for automatic and accurate brain tumor segmentation, 3D visualization, and survival prediction. Our approach consists of three phases: segmentation, feature extraction, and survival prediction. In the first phase, we preprocess the data, perform multi-model fusion, and employ a U-Net architecture for tumor segmentation. The second phase involves extracting latent features from the U-Net encoder and combining them with radiomic features (shape, area, volume, and location) of the segmented tumor subcomponents. The third phase includes feature selection, outlier elimination, bootstrapping, and finally, survival prediction using linear regression. Our results demonstrate a Dice coefficient of 84%, 86%, and 91% for enhancing tumor, tumor core, and whole tumor regions, respectively, on the BraTS2020 dataset. Under post processing, 3D tumor volume is rendered and visualized from segmented results. Average tumor volume size is measured in cm ³ and compared with ground truth volumes. Survival prediction yields an R² score of 0.73, classified into high-risk (0-586 days), moderate-risk (587-1168 days), and low-risk (1169-1767 days) categories, with a random forest accuracy of 85.3%. The proposed method outperforms competitive methods in both tumor substructure segmentation and survival forecast, showcasing its potential for clinical applications.