Aggregating Multi-scale Prediction Based on 3D U-Net in Brain Tumor Segmentation

Abstract

Magnetic resonance imaging (MRI) is the dominant modality used in the initial evaluation of patients with primary brain tumors due to its superior image resolution and high safety profile. Automated segmentation of brain tumors from MRI is critical in the determination of response to therapy. In this paper, we propose a novel method which aggregates multi-scale prediction from 3D U-Net to segment enhancing tumor (ET), whole tumor (WT) and tumor core (TC) from multimodal MRI. Multi-scale prediction is derived from the decoder part of 3D U-Net at different resolutions. The final prediction takes the minimum value of the corresponding pixel from the upsampling multi-scale prediction. Aggregating multi-scale prediction can add constraints to the network which is beneficial for limited data. Additionally, we employ model ensembling strategy to further improve the performance of the proposed network. Finally, we achieve dice scores of 0.7745, 0.8640 and 0.7914, and Hausdorff distances (95th percentile) of 4.2365, 6.9381 and 6.6026 for ET, WT and TC respectively on the test set in BraTS 2019.