Multi-Dilated Hierarchical Filter Based 3D U-Net for Multi-Modal Brain Tumor Segmentation

Brain tumor segmentation is an essential task in the field of medical image analysis. Brain lesions (e.g. gliomas) have various shapes, sizes, and locations between different patients, making it challenging localizing and doing other tasks. For a single tumor, the essential contextual information may inhabit much larger regions than itself. Multi-scale feature representation can improve model generalization. However, existing U-Net models perform with layer-wise features. In this paper, a 3D multi-dilated hierarchical filter module is proposed based on hierarchical multi-scale strategy and multi-fiber strategy. The main points of our proposed module are multi hierarchical filter connection and multi-branch dilated convolutions. Besides, the proposed module is applied with a 3D U-Net. The results of experiments on the validation subset of BraTS2018 dataset show the efficacy of the proposed module.