Multimodal Priors Guided Segmentation of Liver Lesions in MRI Using Mutual Information Based Graph Co-Attention Networks
Multimodal Priors Guided Segmentation of Liver Lesions in MRI Using Mutual Information Based Graph Co-Attention Networks
复制标题
DOI：
10.1007/978-3-030-59719-1_42
复制
发表时间：
2020-10
期刊：
影响因子：
--
通讯作者：
Shaocong Mo;Ming Cai;Lanfen Lin;Ruofeng Tong;Qingqing Chen;F. Wang;Hongjie Hu;Yutaro Iwamoto;Xianhua Han;Yenwei Chen
中科院分区：
文献类型：
--
作者：
Shaocong Mo;Ming Cai;Lanfen Lin;Ruofeng Tong;Qingqing Chen;F. Wang;Hongjie Hu;Yutaro Iwamoto;Xianhua Han;Yenwei Chen
肝脏局灶性病变的分割是初步诊断、分期和治疗后疗效评估的重要预处理步骤。多模态 MRI 扫描（例如 T1WI、T2WI）可提供肝脏病变的补充信息，并广泛用于诊断。然而，某些模态（例如T1WI）具有高分辨率，但缺乏其他模态（T2WI）的重要视觉信息（例如边缘），使用其他模态先验（T2WI）增强T1WI中的组织病变质量并提高分割性能具有重要意义。在本文中，我们提出了一种基于图学习的方法，其动机是有效提取模态特定特征并有效建立 T1WI 和 T2WI 之间的区域对应关系。我们首先将深层特征投影到图域中，并利用图卷积在所有区域中传播信息，以提取特定于模态的特征。然后，我们提出了一种基于互信息的图共同关注模块来学习一个二分图的权重系数，该二分图是通过图域中不同模态的图的全连接构造的。最后，我们通过重投影和残差连接得到最终的细化特征进行分割。我们在多模态 MRI 肝脏病变数据集上验证了我们的方法。实验结果表明，与现有方法相比，该方法通过学习多模态先验（T2WI）的引导特征，实现了 T1WI 肝脏病灶分割的改进。
Segmentation of focal liver lesions serves as an essential preprocessing step for initial diagnosis, stage differentiation, and post-treatment efficacy evaluation. Multimodal MRI scans (e.g., T1WI, T2WI) provide complementary information on liver lesions and is widely used for diagnosis. However, some modalities (e.g., T1WI) have high resolution but lack of important visual information (e.g., edge) belonged to other modalities (T2WI), it is significant to enhance tissue lesion quality in T1WI using other modality priors (T2WI) and improve segmentation performance. In this paper, we propose a graph learning based approach with the motivation of extracting modality-specific features efficiently and establishing the regional correspondence effectively between T1WI and T2WI. We first project deep features into a graph domain and employ graph convolution to propagate information across all regions for extraction of modality-specific features. Then we propose a mutual information based graph co-attention module to learn weight coefficients of one bipartite graph, which is constructed by the fully-connection of graphs with different modalities in the graph domain. At last, we get the final refined features for segmentation by re-projection and residual connection. We validate our method on a multimodal MRI liver lesion dataset. Experimental results show that the proposed approach achieves improvement of liver lesion segmentation in T1WI by learning guided features from multimodal priors (T2WI) compared to existing methods.