基于空间多层次自监督深度学习网络的质谱成像超分辨重建方法研究  

Research on super-resolution reconstruction of mass spectrometry imaging using spatially multi-level and self-supervised deep learning network

作　　者：林超隆 杨慧 葛雅辉 LIN Chao-long;YANG Hui;GE Ya-hui(Institute of Medical Technology,Peking University Health Science Center,Beijing 100083,China;Translational Cancer Research Center,Peking University First Hospital,Beijing 100034,China;Institute of Advanced Clinical Medicine,Peking University,Beijing 100191,China)

机构地区：[1]北京大学医学部医学技术研究院,北京100083 [2]北京大学第一医院肿瘤转化研究中心,北京100034 [3]北京大学临床医学高等研究院,北京100191

出　　处：《医疗卫生装备》2025年第4期1-8,共8页Chinese Medical Equipment Journal

基　　金：国家资助博士后研究人员计划(GZC20230169)。

摘　　要：目的:为了提高质谱成像的分辨率,提出一种基于空间多层次自监督深度学习网络的质谱成像超分辨重建方法。方法:首先,基于非线性变换将组织学图像和质谱图像进行配准;其次,利用多分支视觉变换器(vision transformer,ViT)以自监督学习的方式提取高分辨率组织学图像的层次化特征;最后,将这些特征与配对的低分辨率质谱信息结合构建回归网络,从而实现高分辨率质谱信息的预测。为验证提出方法的性能,与传统基于插值处理的BI(bicubic interpolation)方法和基于卷积神经网络的deepFERE方法对人类肝癌样本金属质谱镁元素图像超分辨重建的结果进行对比,并将其应用于小鼠肾腺癌代谢物质谱成像数据集。结果:与BI方法和deepFERE方法相比,提出的方法在重建质谱图像时表现出最低的均方根误差(root mean square error,RMSE)(RMSE=0.015)、最高的结构相似性指数(structural similarity index measure,SSIM)(SSIM=0.84)和最高的线性回归相关系数(R^(2)=0.853)。通过对小鼠肾腺癌代谢物质谱成像数据集超分辨重建验证了提出的方法的有效性以及精准区分组织特异性的潜力。结论:相较于传统的单模态和逐像素点回归的深度学习方法,提出的方法提高了质谱图像高分辨重建的质量,可作为质谱成像领域超分辨重建的新方法。Objective To propose a method for mass spectrometry imaging(MSI)super-resolution reconstruction based on a spatially multi-level and self-supervised deep learning network(SMSDL-Net),aiming to improve the resolution of mass spectrometry images.Methods SMSDL-Net firstly registered histological and mass spectrometry images using a nonlinear transformation.Then a multi-branch Vision Transformer(ViT)was utilized to extract hierarchical features of highresolution histological images in a self-supervised manner.These features were subsequently combined with the paired low-resolution mass spectrometry data to construct a regression network,which could realize the prediction of highresolution mass spectrometry information.To validate the performance of the proposed method,the results by the method were compared with those of the traditional bicubic interpolation(BI)methods based on interpolation processing and the deepFERE method based on convolution neural network(CNN)for super-resolution reconstruction of magnesium elemental image of metal mass spectrometry of human liver cancer samples,and the method was also applied to a mouse renal adenocarcinoma metabolite mass spectrometry imaging dataset.Results Compared with the traditional BI methods and the deepFERE multimodal method,the method proposed demonstrated the lowest root mean square error(RMSE=0.015),the highest structural similarity index measure(SSIM=0.84)and the highest R-squared value(R^(2)=0.853)in reconstructing mass spectrometry images.The effectiveness of the method and its potential for precise tissue-specific distinction were validated using the mouse renal adenocarcinoma metabolite MSI dataset.Conclusion Compared with traditional singlemodal and pixel-wise regression deep learning methods,the method proposed enhances the quality of high-resolution mass spectrometry image reconstruction and can serve as a novel method for super-resolution reconstruction in the field of mass spectrometry imaging.

关 键 词：质谱成像 超分辨重建 深度学习 多模态图像融合 自监督学习 

分 类 号：R318[医药卫生—生物医学工程] TP183[医药卫生—基础医学]