Joint analysis of multimodalities in spatial mass spectrometry imaging(SMSI)data,including histology,spatial location,and molecule data,allows us to gain novel insights into tissue structures.However,the significant d...Joint analysis of multimodalities in spatial mass spectrometry imaging(SMSI)data,including histology,spatial location,and molecule data,allows us to gain novel insights into tissue structures.However,the significant differences in characteristics such as scale and heterogeneity among the multimodal data,coupled with the high noise levels and uneven quality of MSI data,severely hinder their comprehensive analysis.Here,we introduce a cross-graph cycle attention model,MSCG,to learn efficient joint embeddings for multimodalities of SMSI data by integrating graph attention autoencoders and attention-transfer.Specifically,MSCG enables leveraging one modality(e.g.,histology)to fine-tune the graph neural network trained for another modality(e.g.,MSI).Our study on real datasets from different platforms highlights the superior capacities of MSCG in dissecting cellular heterogeneity,as well as in denoising and aggregating MSI data.Notably,MSCG demonstrates versatile applicability across MSI data from various platforms,showcasing its potential for broad utility in this field.展开更多
随着互联网中多模态数据的快速增长,跨模态检索技术受到了广泛关注。然而,现实中一些多模态数据存在语义信息缺失,导致模型难以准确提取出其中蕴涵的语义特征。此外,一些多模态数据还包含了与语义无关的冗余信息,干扰了模型对关键信息...随着互联网中多模态数据的快速增长,跨模态检索技术受到了广泛关注。然而,现实中一些多模态数据存在语义信息缺失,导致模型难以准确提取出其中蕴涵的语义特征。此外,一些多模态数据还包含了与语义无关的冗余信息,干扰了模型对关键信息的提取。为此,提出了一种基于信息互补与交叉注意力(ICCA)的跨模态检索方法。该方法利用图卷积网络(GCN)建模多标签和数据之间的关系,以补充多模态数据中缺失的语义信息与多标签中缺失的样本细节信息。此外,交叉注意力子模块利用多标签信息,过滤掉数据中语义无关的冗余信息。为了使语义相似的图像和文本在公共表示空间中实现更好的匹配,还提出了一种语义匹配损失。此损失将多标签嵌入融入到图像和文本的匹配过程中,用于进一步增强公共表示的语义性。在NUS-WIDE、MIRFlickr-25K和MS-COCO这三个广泛使用的数据集上进行实验,实验结果表明,ICCA在这些数据集上的平均精度均值(mean average precision,mAP)分别为0.808、0.859和0.837,显著优于现有方法。展开更多
基金supported by the National Natural Science Foundation of China under Grant No.32300523the Shanghai Sailing Program under Grant No.22YF1401700+1 种基金the Fundamental Research Funds for the Central Universities of China under Grant No.2232022Dthe Shanghai Science and Technology Program under Grant No.20DZ2251400.
文摘Joint analysis of multimodalities in spatial mass spectrometry imaging(SMSI)data,including histology,spatial location,and molecule data,allows us to gain novel insights into tissue structures.However,the significant differences in characteristics such as scale and heterogeneity among the multimodal data,coupled with the high noise levels and uneven quality of MSI data,severely hinder their comprehensive analysis.Here,we introduce a cross-graph cycle attention model,MSCG,to learn efficient joint embeddings for multimodalities of SMSI data by integrating graph attention autoencoders and attention-transfer.Specifically,MSCG enables leveraging one modality(e.g.,histology)to fine-tune the graph neural network trained for another modality(e.g.,MSI).Our study on real datasets from different platforms highlights the superior capacities of MSCG in dissecting cellular heterogeneity,as well as in denoising and aggregating MSI data.Notably,MSCG demonstrates versatile applicability across MSI data from various platforms,showcasing its potential for broad utility in this field.
文摘随着互联网中多模态数据的快速增长,跨模态检索技术受到了广泛关注。然而,现实中一些多模态数据存在语义信息缺失,导致模型难以准确提取出其中蕴涵的语义特征。此外,一些多模态数据还包含了与语义无关的冗余信息,干扰了模型对关键信息的提取。为此,提出了一种基于信息互补与交叉注意力(ICCA)的跨模态检索方法。该方法利用图卷积网络(GCN)建模多标签和数据之间的关系,以补充多模态数据中缺失的语义信息与多标签中缺失的样本细节信息。此外,交叉注意力子模块利用多标签信息,过滤掉数据中语义无关的冗余信息。为了使语义相似的图像和文本在公共表示空间中实现更好的匹配,还提出了一种语义匹配损失。此损失将多标签嵌入融入到图像和文本的匹配过程中,用于进一步增强公共表示的语义性。在NUS-WIDE、MIRFlickr-25K和MS-COCO这三个广泛使用的数据集上进行实验,实验结果表明,ICCA在这些数据集上的平均精度均值(mean average precision,mAP)分别为0.808、0.859和0.837,显著优于现有方法。
