Brain Tumor Segmentation Based on 3D Residual U-Net

We propose a deep learning based approach for automatic brain tumor segmentation utilizing a three-dimensional U-Net extended by residual connections. In this work, we did not incorporate architectural modifications to the existing 3D U-Net, but rather evaluated different training strategies for potential improvement of performance. Our model was trained on the dataset of the International Brain Tumor Segmentation (BraTS) challenge 2019 that comprise multi-parametric magnetic resonance imaging (mpMRI) scans from 335 patients diagnosed with a glial tumor. Furthermore, our model was evaluated on the BraTS 2019 independent validation data that consisted of another 125 brain tumor mpMRI scans. The results that our 3D Residual U-Net obtained on the BraTS 2019 test data are Mean Dice scores of 0.697, 0.828, 0.772 and Hausdorff\\(_{95}\\) distances of 25.56, 14.64, 26.69 for enhancing tumor, whole tumor, and tumor core, respectively.