Multi-branch Learning Framework with Different Receptive Fields Ensemble for Brain Tumor Segmentation

摘要

Segmentation of brain tumors from 3D magnetic resonance images (MRIs) is one of key elements for diagnosis and treatment. Most segmentation methods depend on manual segmentation which is time consuming and subjective. In this paper, we propose a robust method for automatic segmentation of brain tumors image, the complementarity between models and training programs with different structures was fully exploited. Due to significant size difference among brain tumors, the model with single receptive field is not robust. To solve this problem, we propose our own method: i) a cascade model with a 3D U-Net like architecture which provides small receptive field focus on local details, ii) a 3D U-Net model combines VAE module which provides large receptive field focus on global information, iii) redesigned Multi-Branch Network with Cascade Attention Network, which provides different receptive field for different types of brain tumors, this allows to scale differences between various brain tumors and make full use of the prior knowledge of the task. The ensemble of all these models further improves the overall performance on the BraTS2019 [10] image segmentation. We evaluate the proposed methods on the validation DataSet of the BraTS2019 segmentation challenge and achieved dice coefficients of 0.91, 0.83 and 0.79 for the whole tumor, tumor core and enhanced tumor core respectively. Our experiments indicate that the proposed methods have a promising potential in the field of brain tumor segmentation.