3D U-Net-Norm architecture for improving generalization of BraTS images

In this work, we augment the 3D U-Net architecture with state-of-the-art (SOTA) generalization techniques using proper parameterization. We empirically assess the normalization and regularization technique(s) with a standard 3D U-Net architecture for glioma segmentation: Edema, Non-enhancing, and Enhancing tumors. We design a fine-tuned normalized 3D U-Net-Norm architecture. We empirically show that we can obtain superior and stable results with proper parametrization of the combination of SOTA techniques. We empirically assess the proposed model using the combined – BraTS-2020 and BraTS-16-17 – 3D MRI datasets. The simulation findings in training and validation loss and other performance measures show that the proposed 3D U-Net-Norm gives a more stable performance than other techniques across multiple runs on all BraTS data and, thus, tends to provide a good generalized architecture for brain tumor segmentation. The proposed work is a step toward increasing the trust of a radiologist in DL-based assisted technology.