Low-Rank Convolutional Networks for Brain Tumor Segmentation

The automated segmentation of brain tumors is crucial for various clinical purposes from diagnosis to treatment planning follow-up evaluations. vast majority effective models tumor are based on convolutional neural networks with millions parameters being trained. Such complex can be highly prone overfitting especially in cases where the amount training data insufficient. In this work, we devise a 3D U-Net-style architecture residual blocks, which low-rank constraints imposed weights layers order reduce overfitting. Within same architecture, helps design several times fewer parameters. We investigate effectiveness proposed technique BraTS 2020 challenge.