Quality-Aware Model Ensemble for Brain Tumor Segmentation

Automatic segmentation of brain tumors is still a challenging task. To improve the segmentation performance and better ensemble all the candidate models with different architectures, we proposed a three-stage model with the quality-aware model ensemble. The first stage locates the tumor with coarse segmentation, while the second stage refines the coarse segmentation in the region of interest. The last stage performs the quality-aware model ensemble with a quality score prediction net to fuse the results from the multiple outputs of sub-networks. Besides, we warp a standard SRI24 brain template to the subject image, which is a strong prior of the brain structure and symmetry. Our method shows competitive performance on the BraTS 2021 online validation dataset, obtaining an average dice similarity coefficient (DSC) of 0.911, 0.850, 0.816, and average 95th percentile of Hausdorff distance (HD95) of 4.58, 8.959, 10.400, for whole tumor, tumor core, and enhancing tumor, respectively.