Pushing back the limits in the analysis of mass spectrometryimagesof biological tissues by using parallel factor analysis

For much of the past decade, hyperspectralimaging (HSI) has been an area of active research and development which has been introduced in many applications in chemistry, medicine, agriculture, mineral exploration, and environmental monitoring.Mass spectrometry imaging (MSI)is one of the most powerful HSI techniques that extends the capability of traditional imaging techniques by obtaining spatial images of a sample at a series of ( 100) continuousmass to charge ratios (m/z)[1].A MSIimage is a three-dimensional (3D) hyperspectralcube which is composed of vector pixels containing spectral information (of m/z values) as well as twodimensional spatial information (of xrows and ycolumns).Due to the complexity of MS images, chemometric methods have shown potential forthe analysis of MSI data [2]. In the present contribution, a chemomtricstrategy based on binning approach for image compression and parallel factor analysis (PARAFAC) [3] for image resolution was developed for analysis of MS images obtained from different sections ofmouse lung.Due to the huge size of MS image of mouse lung sections(10 billion elements of 49×132 pixels and 500000 m/zvalues), images were compressed using binning approach in m/z direction (bin size=0.275) as the most common compression way. In this way, the number of elements was reduced from 10000 million elements to 1.2 million ones (49×132×6000)which is approximately 1% of the original size. Then, the number of components in the analyzed section was determined using core consistency diagnostics (CORCONDIA)which was 10 in this case. Then, the trilinear model assumption of the data was tested using singular value decomposition (SVD) of the row-and column-wise augmented data. The results confirmed the trilinearity of the image cubes. Therefore, PARAFAC was used for image resolution with GRAM/DTLD as initial estimates to start ALS optimization. Also,non-negativity constraintwas applied tothreedatamodes[3]. Using this method, both spatial distribution and spectral information of analyzed samples wereobtainedwith lack of fit (LOF) values below 10%.The 2D distribution maps for different components were then obtained using a post-processing step. Evaluation of the results showedthree different lung regions based on differences in resolved mass spectra anddistribution maps.These three regions were related to external membrane of the lung, parenchyma region and blood vessels in the lung. Finally, the results of this study were compared with multivariate curve resolution-alternating least squares (MCR-ALS) which was the aim of previous studies [2]. Inspection of the results showed the equivalence of the results of both methods. It is concluded that combination ofMSIand advanced data analysis tools suchas PARAFAC has allowed the extraction of valuableinformation from a highly complex massive datasetlike mouse lung.