文献类型：期刊文献

英文题名：Mutual ensemble learning for brain tumor segmentation

作者：Hu, Jingyu[1];Gu, Xiaojing[1];Gu, Xingsheng[1]

机构：[1]East China Univ Sci & Technol, Key Lab Smart Mfg Energy Chem Proc, Minist Educ, Shanghai 200237, Peoples R China

年份：2022

卷号：504

起止页码：68

外文期刊名：NEUROCOMPUTING

收录：;EI(收录号：20222912384974);WOS:【SCI-EXPANDED(收录号:WOS:000861317600006)】；

基金：This work is supported by National Natural Science Foundation of China under Grant No. 61973122 and 61973120.

语种：英文

外文关键词：Mutual ensemble learning; Model ensemble; Convolutional network; Brain tumor; Segmentation

摘要：It is challenging to reduce the generalization errors of brain tumor segmentation models on test data, as the nature of the high diversity of tumors. The model ensemble combining multiple models to make the final prediction is a reliable strategy to reduce generalization errors. Conventionally, these models that acted as member networks of the ensemble are trained separately. Each member network is trained inde-pendently on a data subset or a single perspective of 3D brain volumes. To enable each member network to be trained on the complete training dataset as well as 3D images, we propose a mutual learning method to mutually train multiple 3D segmentation networks of the ensemble, called mutual ensemble learning (MEL). Mutual learning can enable knowledge exchange between networks and let them teach each other during the training process so that each member can converge to a better local minimum com-pared with when trained separately. Meanwhile, mutual learning also keeps the partially independent errors made by different member networks, and combining these member networks can reduce overall errors on the test set. In a word, combining these stronger members gives better ensemble results and it would not bring more cost in the test stage. To realize the mutual learning of multiple 3D brain tumor segmentation networks, we introduce a novel loss called Consensus Dice loss to explicitly exchange infor-mation among different networks during training. This loss considers the overlap of prediction probabil-ities by multiple networks and ground truth to enhance the overlapping pixels' probabilities. Extensive experiments are conducted on the brain tumor segmentation, i.e., BRATS 2015 and 2018 datasets. Results on the databases indicate that the proposed approach consistently improves the segmentation performance of the baseline network. Furthermore, our proposed method achieves state-of-the-art per-formance on the BRATS 2018 online validation set, and it is flexible to be extended by diverse homoge-neous and heterogeneous state-of-the-art networks.(c) 2022 Elsevier B.V. All rights reserved.