Multimodal Brain Tumor Segmentation Using Modified 3D UNet3+ Architecture

Glioma is a type of tumor that originates in the glial cells of the brain or the spinal cord, which is the most common and aggressive malignant primary tumor of the central nervous system in adults. Magnetic resonance imaging (MRI) is an efficient way of detection for brain tumors, it contains precise knowledge about the location of the tumor and its component. Manual tumor labeling is a tedious and time-consuming task, hence accurate automatic segmentation has important clinical significance in the diagnosis, and deep learning shows great power on it. In this paper, we propose an improved 3D U-Net3 + segmentation network. We adopt the residual structure in the encoder part to address the challenge of diminishing gradient problem, as well as to enhance the encoder’s capability for feature extraction, thereby facilitating comprehensive feature fusion during up-sampling within the network. Meanwhile, we add a critic network at the end of the U-Net3+, that contributes to the generation of segmentation results that are more authentic and dependable from the segmentation network. We trained and evaluated the architecture on the BraTS 2023 dataset, and achieved LesionWise_Dice of 88.22%, 83.85%, 81.97% for the whole tumor, tumor core and enhancing tumor, respectively.