基于人工神经网络的生物组织质谱成像分类与识别方法

Artificial Neural Networks Method of Classification and Identification for Mass Spectrometry Imaging Data of Biological Tissue

生物组织质谱成像技术不仅能够展示组织的生物分子信息,而且能直观地显示分子空间分布,是当今生物质谱的研究热点。如何对生物组织质谱成像的数据进行基于生物分子的有效分类与识别是该领域关注的重要问题,特别对于病变组织与其邻近非病变组织的区分与识别和生物组织功能区域的划分与鉴定具有重要的意义。本研究开发出一种新的分类与识别方法。其流程是,首先进行质谱成像数据预处理,应用无监督的自组织特征映射网络区分组织样品区与非组织区域,提取组织区域的质谱数据,应用有监督的学习向量量化网络对已知类别数据进行学习训练,建立模型;应用模型对未知样品进行识别。应用本方法对6个膀胱癌患者的膀胱癌变组织与邻近非癌变组织的质谱成像数据进行分类与识别,结果显示,癌变组织判错率低于23.38%,而非癌变组织判错率低于9.08%,表现出较高的准确度;对3片邻近的小鼠大脑切片质谱成像数据进行白质与灰质区域划分,将中间的1片用于训练,两边的2片用于验证,结果显示,自组织特征映射网络的分类结果与学习向量量化网络的预测结果不一致率低于4%。本方法基于生物分子的质谱成像组织区域分类与识别,具有较高准确度和操作简便等优点,在临床医学研究领域有大规模的应用潜能。

Mass spectrometry imaging（MSI）,the combination of molecular mass analysis and spatial information,provides visualization of molecules on complex biological surfaces,thus is currently getting a significant amount of attention in the mass spectrometric community.One important problem in this researching field is how to develop an effective method of classification and identification for MSI data,especial for identifying the cancerous tissue from adjacent normal tissue and classifying the different functional regions in a complex biological tissue.For this purpose,we developed a new method,containing image reconstruction from raw mass spectral data,MSI data pre-processing,classification of tissue regions from background regions by self-organizing feature map and identification of special interesting regions from the whole tissue regions by learning vector quantization.The MSI data of six pairs（12 tissue samples） of human cancerous and adjacent normal bladder tissue samples were used to test the effect of this method.The result showed an error rate of less than 23.38% for identification of cancerous regions and an error rate of less than 9.08% for identification of the adjacent normal regions.The method was also tested to classify white matter and gray matter regions of three adjacent slices of mouse brain tissue.The slice in the middle was used to train and to establish an identification model;the other two slices were used to test the model.The inconsistent rate of the identification results by using self-organizing feature map is less than 4% comparing with the results using learning vector quantization.This indicated that the method could be performed simply and efficiently,to extend the capability of MSI,and underline its potential to be a regular tool applied to study on clinical application.