基于同步辐射X射线荧光光谱与一维卷积神经网络的癌症筛查方法

Research on Cancer Screening Method Based on Synchrotron Radiation X-ray Fluorescence Spectroscopy and One-dimensional Convolutional Neural Network

癌症是全球范围内引起高发病率与高死亡率的疾病之一。现有癌症检测方法耗时、昂贵、专业人员依赖性强,开发一种无损、快速筛查方法非常重要。在前期工作基础上,发展了基于同步辐射X射线荧光光谱技术(SRXRF)与深度学习技术结合的一种非靶标金属组学方法筛查癌症患者。首先,分析控制组与癌症组共269份血清样本的SRXRF谱线,得到Ca、Mn、Zn、Ge、Br在两类人群中具有代表性差异,可以作为癌症筛查的标志物;其次,对于平均光谱进行归一化(Normalization)、迭代自适应加权惩罚最小二乘法(airPLS)、Savitzky-Golay平滑(SG)、标准正态变换(SNV)的预处理,并建立偏最小二乘判别分析(PLSDA)、K近邻法(KNN)、软独立建模分类法(SIMCA)的化学计量学模型,三种模型对癌症筛查的最优准确率分别为89.89%、93.26%、90.95%;最后,基于像素级光谱,搭建三种一维卷积神经网络(1DCNN)模型,三种模型准确率分别为93.56%、95.24%、93.27%,相对于化学计量学模型均有所提高,增加卷积层的数量有助于数据特征提取,模型准确率提高了1.68%。将三种模型卷积层提取获得的特征进行t-分布随机邻域嵌入算法(tSNE)降维可视化,得到1DCNN提取的特征具有显著可分性,SRXRF结合1DCNN模型开发的非靶标金属组学方法在实现癌症的快速筛查方面具有潜力。

Cancer is one of the diseases that cause high incidence rate and mortality worldwide.Existing cancer detection methods are time-consuming,expensive,and highly dependent on professionals,making it crucial to develop a non-destructive and rapid screening method.On the basis of previous work,this article developed a non target metabolomics method for screening cancer patients based on the combination of Synchrotron Radiation X-ray Fluorescence Spectroscopy(SRXRF)and deep learning technology.Firstly,by analyzing the SRXRF spectra of 269 serum samples from the control group and the cancer group,it was found that Ca,Mn,Zn,Ge,and Br had representative differences between the two populations and could be used as biomarkers for cancer screening.Secondly,normalization,adaptive iteratively reweighted penalized least squares(airPLS),Savitzky-Golay smoothing(SG),and Standard Normal Variate(SNV)were performed on the average spectrum,and chemometric models of Partial Least Squares Linear Discriminant Analysis(PLSDA),K-Nearest Neighbor(KNN),and Soft Independent Modeling of Class Analogy(SIMCA)were established.The optimal accuracy rates of the three models for cancer screening were 89.89%,93.26%,and 90.95%,respectively.Finally,based on pixel level spectra,three one-dimensional Convolutional Neural Network(1DCNN)models were constructed,with accuracy rates of 93.56%,95.24%and 93.27%,respectively.Compared with chemometric models,all models improved significantly.Increasing the number of convolutional layers was conducive to extract data features,and the model accuracy improved by 1.68%.The features extracted from the convolutional layers of three models were visualized using t-distributed Stochastic Neighbor Embedding(tSNE)for dimensionality reduction,and it was found that the features extracted by 1DCNN had significant separability.In summary,the nontargeted metallomics method developed by SRXRF combined with 1DCNN model is potential in achieving rapid cancer screening.