J. Phys. IVFrance 104 (2003) 617 @ EDP Sciences, Les Ulis DOI : 10. 1051/jp4 : 20030156 Data analysis for X-ray fluorescence imaging S. Vogt, J. Maser and C. Jacobsen1 Expen'menfa/FacMes D ; s ; bn, rgonne Naf/ona/Laborafory, 9700 South Cass Avenue, Argonne, IL 60439, U. S. A. 1 Department of Physics and Astronomy, state University of New York at stony Brook, Stony Brook, U. S. A. Abstract. X-ray-microprobe-based X-ray fluorescence (XRF) scanning microscopy is a powerfultechnique to map and quantifyelement distributions in biological specimens, such as cells and bacteria. Principal component analysis (PCA)provides a method to correlate an XRF data set with full spectra at each scan point and to weigheach component of the spectrum, and its corresponding eigenimage, according to its respective significance in the data set. In particular, photon noise is not correlated among pixels and therefore does not contribute to the principal components. We showthat, by fitting the eigenspectra of the principal components, one can then generate maps of fitted elemental components with high accuracy, without the need to fit the spectra of single pixels. Additionally,the correlation of elemental distributions can be used to revealinformation aboutthenumberand composition of the différentmajor constituents of a cell. We also demonstrate that cluster analysis can be used to classify the sample into spatially separate régions of characteristic elemental compositions, for example nucleus, cytoplasm, and vesicles. 1. INTRODUCTION The cellular and subcellular distribution of biologically relevant elements, such as phosphorus (ATP, DNA), zinc (transcription factors), or calcium (second messenger proteins) is a scientific question of strongly increasing importance in biomédical and life sciences research. To answer this question, several différent techniques have been developed. In light microscopy. fluorescent dyes, which change excitation or émission characteristics upon binding to a spécifie chemical element, have been used to visualize a sélect number of métal ions (see, e. g., [1, sect. 20]). In électron probe X-ray microanalysis, scanning électron microscopes with energy-dispersive X-ray detectors are employed to detect the X-ray fluorescence emanating from the small spot of the sample that is probed with the électron beam (see, e.g., [2]). Numerous chemical elements can be detected simultaneously, but bremsstrahlung emitted during incohérent scattering of the électrons leads to high background onto which the actual XRF peaks are superimposed. Also. high spatial resolution requires sectioning of the samples to below 200 nm section thickness. X-ray probe X-ray microanalysis is a powerful technique that we use to map and quantify element distributions in biological specimens, such as cells. High-resolution X-ray optics (e. g., zone plates) are used to focus X-rays into a small spot on the sample. The énergies of the X-ray fluorescence photons emitted from the probed spot are characteristic for the respective elements. The sample is raster-scanned through the focal spot. Rather than recording the signal only within a limited number of predetermined energy Windows, we record the full spectrum of the emitted x-ray fluorescence at each scan point to make possible the analyses described below. The resulting three-dimensional data sets (sample x position, sample y position, photon energy E) are limited in energy resolution tE by existing solid-state detectors (e. g., AE=160 eV at E=5. 8 keV), so that care must be taken in the analysis of fluorescence signais that are nearby each other in energy. Furthermore, in order to realize reasonably short data acquisition times for large scans, the intégration time at each pixel is kept low, which leads to single-pixel spectra with low photon statistics. A common approach is then to reduce the data set to maps representing elements of interest (e. g., Ça) by integrating over spectral régions of interest (ROI filtering). The background can be estimated using dedicated ROIs, and the influence of the Kp lines of elements with high concentration on the ROI of a neighboring element can be either calculated or measured with standard référence material. and then removed from the elemental maps (p-stripping). This has the disadvantage of an increase in noise.