Abstract: The basic procedure of spectral image analysis is mapping, i.e., generation of conventional images using specific spectral features from multidimensional data set. Commonly, it is based on decomposition of the original spectra into a sum of model spectra which are usually attributed to chemically/physically different components. However, the problem is complicated when the model spectra are not clearly definable or difficult and even impossible to obtain experimentally. Here we describe a method for estimating spectral models based on the elimination of physical meaningless solutions obtained by PCA (Principal Component Analysis). The method includes the following steps: segmentation of the original spectral image using a white- light image; generation of primary factors for each segmented region by MDF (maximum distance factors) algorithm; extending the factor space to satisfy non-negative score requirements; generation of model limits by rotation of factor matrix in conditions of non-negative intensity values; calculation of the common model as intersection regions in the score space. The precision of the method depends on the spectral variations in the original data. When these variations are significant the models could be obtained precisely and then attributed to possible sample components. This method was validated in different research fields: to study the distribution of anticancer drugs in single living cells, to characterize phenolic species in wheat walls. !8
展开▼
