文献类型：期刊文献

英文题名：MDANet: Multimodal difference aware network for brain stroke segmentation

作者：Zhang, Kezhi[1];Zhu, Yu[1];Li, Hangyu[1];Zeng, Zeyan[2];Liu, Yatong[1];Zhang, Yuhao[2,3,4]

机构：[1]East China Univ Sci & Technol, Sch Informat Sci & Engn, Shanghai 200237, Peoples R China;[2]Fudan Univ, Zhongshan Hosp, Dept Neurol, Shanghai 200032, Peoples R China;[3]Natl Clin Res Ctr Intervent Med, Shanghai 200032, Peoples R China;[4]Shanghai Clin Res Ctr Intervent Med, Shanghai 200032, Peoples R China

年份：2024

卷号：95

外文期刊名：BIOMEDICAL SIGNAL PROCESSING AND CONTROL

收录：;EI(收录号：20241916052489);WOS:【SCI-EXPANDED(收录号:WOS:001239563800001)】；

基金：Acknowledgment This work was supported by the Exploratory Device R&D Projects of National Clinical Research Center for Interventional Medicine, China (NO. 2021-002) and the Key Clinical Research Projects of National Clinical Research Center for Interventional Medicine, China (NO. 2019-003) .

语种：英文

外文关键词：Computer-aided diagnosis; Brain stroke; Medical image segmentation; Deep learning

摘要：Stroke segmentation has great significance for clinical diagnosis and timely treatment. Medical images of strokes often come in the form of multiple modalities. But most existing methods simply stack these modalities as input, disregarding the connections and other clinical prior knowledge associated with each modality. In this paper, we present MDANet, a multimodal difference aware network for stroke segmentation based on multimodal input. The proposed network mainly consists of a difference aware module and a graph convolution fusion block. In the difference aware module, a parameter -shared encoder is adopted to extract features from different modality groups and generate difference feature maps by subtracting one group from another to enhance the perception of potential lesion areas. We further design a similarity loss to improve this ability. The graph convolution fusion block is developed to aggregate features from different modalities with a channel embedding strategy to model the features globally and a space embedding strategy for local modeling. The MDANet is trained and evaluated on the Ischemic Stroke Lesion Segmentation (ISLES) 2018 and 2022 datasets. Our approach achieves a dice score of 58.34 and 70.44, surpassing the performance of other advanced existing methods.