Segmentation and Classification of Brain Tumor Using 3D-UNet Deep Neural Network

Abstract

Brain tumor segmentation and classification are critical steps in neuro-oncology for treatment planning and prognosis assessment. This paper proposes a 3D-UNet deep neural network framework for simultaneous brain tumor segmentation and multi-class classification from multimodal MRI scans. The proposed architecture processes T1, T1ce, T2, and FLAIR MRI sequences as multi-channel 3D volumes, leveraging residual connections and channel attention modules to improve feature discrimination between tumor sub-regions. The system segments tumors into whole tumor, tumor core, and enhancing tumor regions and simultaneously classifies tumor grade (LGG vs. HGG) through a shared encoder with task-specific decoders. The model achieves Dice scores of 0.89 (whole tumor), 0.82 (tumor core), and 0.76 (enhancing tumor) on BraTS 2021, with tumor grade classification accuracy of 91.4%, outperforming standard 3D-UNet and V-Net baselines.