STXMPy: a new software package for automated region of interest selection and statistical analysis of XANES data
© Haraszti et al 2010
Received: 29 January 2010
Accepted: 4 June 2010
Published: 4 June 2010
Soft X-ray spectromicroscopy based absorption near-edge structure analysis, is a spectroscopic technique useful for investigating sample composition at a nanoscale of resolution. While the technique holds great promise for analysis of biological samples, current methodologies are challenged by a lack of automatic analysis software e. g. for selection of regions of interest and statistical comparisons of sample variability.
We have implemented a set of functions and scripts in Python to provide a semiautomatic treatment of data obtained using scanning transmission X-ray microscopy. The toolkit includes a novel line-by-line absorption conversion and data filtering automatically identifying image components with significant absorption. Results are provided to the user by direct graphical output to the screen and by output images and data files, including the average and standard deviation of the X-ray absorption spectrum. Using isolated mouse melanosomes as a sample biological tissue, application of STXMPy in analysis of biological tissues is illustrated.
The STXMPy package allows both interactive and automated batch processing of scanning transmission X-ray microscopic data. It is open source, cross platform, and offers rapid script development using the interpreted Python language.
Scanning transmission X-ray microscopy (STXM) is a synchrotron based technique for the investigation of sample structure and composition with nanoscale (c. a. 30 - 50 nm) resolution [1, 2]. High resolution X-ray microscopy is based on X-ray absorption spectroscopy and X-ray absorption near-edge structure analysis (XANES) which provides the chemical information about the specimen.
Compared to electrons soft X-rays have excellent tissue penetrating capability. Using photon energies in the so called "water window" between the carbon and oxygen K-shell absorption edges, STXM allows imaging of naturally occurring absorption contrast differences within biological samples. The spectral information of soft X-ray XANES combined with the high spatial resolution of STXM near the carbon or the oxygen K-shell energy (about 284 eV or about 533 eV) holds promise for discovering and studying chemical changes underlying a wide-range of biological phenomenon and disease states.
One challenge in the biological application of these techniques pertains to sample variability within and between individual preparations. Biological samples tend to be highly heterogeneous. Accordingly, biological applications of STXM and XANES require larger number of analyses in order to perform experiments with statistical significance. Currently, analysis of STXM data is typically completed using software packages such as the one created by the X-ray physics group of the Stony Brook University or the aXis2000 software provided by the McMaster University. Both packages are written in the interactive data language (IDL, Visual Information Solutions) and offer many powerful tools such as automatic stack alignment. Unfortunately, spectral data averaging of both packages is based on image areas selected manually by the user. Thus, neither are ideal for biological samples requiring analysis of many regions of interest and both are subject to potential user bias in selection of regions of interest.
Here, we present a new software package for analyzing STXM data based upon a simplistic analysis approach, and including a line-by-line absorption conversion tool. By automating the selection of regions of interest, the approach empowers analyses of large biological data sets. In developing this software, we analyzed melanosomes, the sub-cellular organelle responsible for melanin pigment production. As expected, the variability within data from melanosomes was found to be very high. However, the high number of data points analyzed through use of the STXMPy [Additional file 1] software package empowered a statistically meaningful analysis to be performed and was able to identify spectral differences between organelles isolated from mice with known genetic differences.
All the programs described below were written in the interpreted language Python, and are based on three main libraries: the NetCDF library pycdf from Unidata, the numpy library  and the matplotlib plotting library . For testing and development the ipython interface was used, which allows command history and history recording .
The basic sm object
In order for data to be interpreted, image data and related parameters must first be integrated. To achieve this, the basic sm object is located in the sm package and acts to load and store a STXM image and stores the energy of the image or the minimum and maximum energy if those are different from one another. The physical extents of the image are calculated and stored, along with all other parameters loaded from the NetCDF file. The snippet below presents an example of picking up an STXM image from a typical experiment recorded using the proportional counter of the X1A STXM. (The 'x1aos_20060412_0162.sm' file is available in the additional files of this paper [Additional file 2 and 3].)
from sm import sm
from matplotlib import pyplot as pl
stxmImage = sm('x1aos_20060412_0162.sm')
imagedata = stxmImage.get_image(-1)
print "Energy: %.2f eV" %stxmImage.E
Reading and aligning stacks
Because STXM based XANES utilizes a stack of STXM images taken at various energies, the achievable spatial resolution can be limited by alignment imperfections. Stack alignment is therefore critical to successful implementation. Based on the xanesP and ImageP packages, the AlignSm.py script presents an application of the available parts, reading and aligning a list of STXM images. For simplicity the list of image files is stored in a single text file one filename per line. This list is read, the files imported, and the images are aligned using a convolution comparison. The program does not modify the sequence of the images, except to invert the sequence if the first energies are in a decreasing order. The script accepts several command line parameters, allowing the entry of arbitrary file name as the list of data files, selection of frames to be used as reference image, and defining a region of interest for the reference image or stack.
For example, to select the text file containing the names of the data files to be 'lst.lst', align the stack relative to image number 50 in this list, and 'despike' the stack before aligning, use:
AlignSM.py -f lst.lst -i 50 -d
The resulting stack will subsequently be displayed on screen and stored on the hard disk together with the list of energies.
Alignment is achieved by the AlignStack() function from the xanesP package. This functionality uses the image convolution built into the ImageP package. This is a standard, fast Fourier transform (FFT) based convolution algorithm [1, 2, 6, 7] with padding (either padding to the added size of the two images as a minimal padding, or alternatively to the next power of 2) . Many biological applications, including our experiments with isolated organelles, will contain relatively few small absorbing features against a large bright background devoid of features. Therefore for the convolution the images are inverted in intensity by default. For images with a large area of absorbing features, inversion can be shut off to achieve a better alignment. If requested, the convolution image is also displayed, so that the user may visually check the result. Because images realigned by subpixel accuracy using a quadrilinear interpolation formula  also apply smoothing, it is advantageous to use pixel level alignment. To achieve this, shifted images were padded by zero for those pixels which were unknown due to the shift, and then the image stack was cropped to the smallest common area containing measurement data. In an ideal case, approximately 1 pixel of uncertainty is expected due to the round off error of the positioning, with greater uncertainty when low contrast or high image noise are present in the images.
Measuring the absorption of materials in X-ray spectroscopy is usually done by recording the transmitted intensity of X-rays through the sample and then trough a blank reference sample providing the background I0 data. This approach, i. e. recording reference measurements before/after the sample, is however not feasible when spectroscopy is derived from STXM image stacks, because recording such stacks takes several hours. Instead, one can envision measuring a reference point before/after each image. This latter scenario is equivalent to recording images with areas void of the sample in interest, providing simultaneous intensity measurement of the supporting substrate and avoiding a high accuracy repositioning of the sample between each step. This approach is generally accepted for STXM stack based spectroscopic analysis and requires the identification the background areas reliably in order to access the background intensity data.
The AbsorptionSm.py script converts the aligned stack data to absorption values and performs normalization (using the NormAStack function from the sm package). The script offers several options, allowing choice between absorption conversion methods (line or image), thresholding values for these methods, and parameters to be passed to the normalization routine. The data are taken from the results of the AlignSm.py and stored again automatically. The script also displays the resulted stack of absorption images. The example below converts the intensity stack to absorption images using a histogram based image conversion and a relative threshold of 0.6. Intensities above 0.6 times the maximum intensity are used to calculate the bright background.
AbsorptionSm.py -a -t 0.6
The script has two implementations of the absorption calculation: an overall conversion and a line-by-line conversion.
In standard spectroscopy the logarithm is often used with a base of 10, but in order to be more consistent with Beer's law, the AbsorptionSm.py script uses the natural base. To automate the process and remove possible user bias, a simple, statistics based approach was used. The background intensity I0 is calculated using the mean value of the higher intensity pixels from the image. These pixels are selected by thresholding the image. By default a dynamic thresholding is used based on Otsu's method and implemented as graythresh() in the ImageP package . The function uses a histogram with a given number of bins (50 by default) to calculate the maximum of the interclass variance. Requesting verbose output will plot the interclass variance data to the screen.
from ImageP import graythresh
from matplotlib import pyplot as pl
th = graythresh(imagedata, bins = 50, verbose = True)
#to display which part of the image is above threshold:
dispimage = imagedata - imagedata.min()
pl.figure() #open a new figure to display the image
pl.imshow(imagedata*(dispimage > th*dispimage.max()),\
The returned value is a floating point number, the relative position of the threshold between the minimum and maximum intensity of the image.
A complicating feature of X-ray absorption spectroscopy based on STXM data is that the intensity of the beam reaching the samples sometimes fluctuates during image recording. The observed fluctuation has two characteristic time scales. In one case fluctuation occurs within single scan lines. This happens only rarely, causing small local perturbations (a few pixel in a line) in the images that are negligible in scale compared to the data of the entire image. More frequently, regional fluctuations occur which change the overall intensity of larger areas in the images. The standard practice in such instances is to drop the image and the given energy value from the image stack. To avoid this loss of data, a line-by-line absorption calculation was introduced.
The algorithm is based on two decisions: 1) selecting lines which contain absorptive features (objects), and 2) selecting pixels to be used for the zero intensity calculation.
To select lines containing features a simple statistics based approach was utilized. The program first analyses each line by taking a simple running average filter, typically with 5 point averaging range. Comparing the standard deviation of the original data to the smoothed data, lines with features must have a smaller reduction in their standard deviation than those containing only fluctuations caused by the beam noise. Lines reaching a critical difference are treated as empty and I0 is calculated as their mean average.
To select pixels for the zero intensity calculation, lines having features are treated to a second manipulation. The goal is to identify an intensity limit above which value the pixels can be taken for calculating the zero intensity. Because each line contains less than 200 data points, histogram based zero intensity determination is not practical. Two alternative approaches were initially tested. In the first approach, all intensities above the mean value of the line were taken into account. In the second approach, only intensities which were in the upper 20% of the intensity range of the data were taken into account. The latter approach proved superior, though there is a potential for "spikes" (single pixels with very high intensity) to corrupt the results. Therefore, the software generates a warning if the data set used to calculate I0 is less than 3 pixels long. The zero intensity (I0) is calculated as the mean value of the selected pixels and then used to convert the whole line to absorption. For the sake of completeness the algorithm is also applicable with various polynomial fittings to reduce known drift within the image lines, as it is a common feature for data processing in atomic force microscopy images and was incorporated for possible use in future instruments.
Traditional STXM data analysis strongly relies on definition of the area of interest (ROI), and therefore the pixels from which the absorption or intensity spectra can be averaged. Averaging of absorption data can be done before  or after normalization.
During normalization, the linear aspect of the pre-edge spectra (below 282.0 eV for the carbon K-shell absorption edge) is frequently subtracted as a background in order to remove the contribution of other elements from the spectrum. The resultant spectra can subsequently be directly fitted with a set Gaussian or Voigt line profiles and a step function to follow the given absorption edge. Normalization can also be achieved either by using this height of the fitted absorption edge or by using the post absorption edge slope of the data [7, 11–13].
Because biological samples are typically of non-uniform thickness and heterogeneous, the preferred methodology employed by the STXMPy package was to perform normalization before averaging. Because high beam noise above 300 eV made the post edge region unaccessible in our experiments with the X1A STXM , the broad absorption peak at 288 eV (aliphatic CH, carboxylic and amide groups ) was chosen as a representative of the organic content of the melanosomes. The peak value was evaluated using a second order polynomial fit in the 287 - 289 eV range. In order to avoid noise amplification with the normalization and background subtraction, each spectrum was first characterized by its mean value and variation. When the characteristic edge step was not found (the spectrum did not increase with at least 0.1 between the normalization range and the background, thus from 280 eV to approximately 288 eV), the image point was considered as a background pixel and its spectrum was left unaltered. The resultant normalized data stack was then saved for later use.
The EvalSm.py program picks up the normalized spectral set resulted by the AbsorptionSm.py, and evaluates the spectra based on comparison to a smooth fit (6th order polynomial). Spectra which have a fitting noise below threshold are averaged, the average and the standard deviation values are stored into a tabulated text file, and plots are generated.
All the above described scripts, which process STXM stacks save their data back to the working folder where they were called from. It is desirable to compare and average these data sets in a flexible way, for which another script is available. The XAS-averager.py accepts a list of folder names in an input file, and creates the average and standard deviation of the spectral data from these folders. The script checks the energy of each data and eliminates errors caused by missing energy values.
Speed of execution
Interpreted programming languages may provide execution speeds inferior to optimized, compiled languages, such as C. To improve work flow performance, the data processing was split into various steps, storing the data between in predefined data files. In this manner, some of the slower steps, such as the alignment of the stack, did not have to be repeated when experimenting with various analysis parameters for absorption conversion or when spectral averaging is executed.
The current, pure Python code has three slow steps: 1) despiking takes approximately 2 minutes for 135 images, which is the slowest step. This could be improved by employing a less general algorithm, which would however limit the broad scope of the script. 2) Alignment, which is based on the already optimized built in FFT from the numpy package, takes 30 - 80 seconds running time. 3) The line-by-line absorption conversion takes approximately 45 seconds, which is acceptable taking into account the simultaneous image feedback provided.
All other processing steps are performed within a few seconds, resulting an acceptable execution speed on a modern personal computer (Pentium class, dual core machine with 2 GB RAM). Increasing the performance of the STXMPy package may be achieved by adding external, compiled alternatives of the various functions. This step may improve the running speed up to two times, depending on the complexity of the algorithm, and may present a part of future development.
Results and Discussion
Imaging of highly purified and freeze dried melanosomes was carried out as described in reference  at the carbon K-shell energies in a He jet atmosphere using the STXM of the X1A beamline of the National Synchrotron Light Source (NSLS), Brookhaven National Laboratories, Upton, NY USA) . Energies 280 - 310 eV were used with up to 0.1 eV resolution.
Comparing the newly introduced line absorption algorithm to its classical counterpart, we were interested in evaluating two aspects of its performance: 1) how the algorithm handles energy shifts within an image and 2) whether the algorithm provides similar results to those of the classical absorption conversion.
These results indicate that the overall output, the obtained XANES spectrum, is robust against various thresholding methods. Furthermore, the absorption conversion line-by-line performs similarly to the classical, whole image conversion method (compare dynamic thresholding to line-absorption in Figure 3c,d).
Here, we have presented a semiautomated data processing method for extracting XANES data from STXM data sets. The analysis allowed average information of samples to be analyzed and statistically compared. Separate programs were written in the Python programming language to process and visualize data sets both in interactive and batch processing. The system includes a line-by-line conversion of images to absorption data and outputs images that can be evaluated as data and quality controls. This algorithm is envisioned to be generally useful in XANES analysis of biological samples and any STXM system where external noise source contributes variations in the intensity of the incoming beam or where automated region of interest selection would be desired.
Availability and requirements
STXMPy: image manipulation for STXM images
Project home page
Linux, but all implementations are platform independent
Python 2.4 or higher (not tested with Python 3)
Python imaging library (not used in this work, but required by some parts of ImageP)
Numerical Python: numpy for array handling and math
Matlab like plotting library: matplotlib for data visualization
NetCDF wrapper from Unidata: pycdf for accessing the sm file format from the X1A beam line
GNU LGPL 3
The authors would like to thank to the National Synchrotron Light Source for the beam time and Dr. Sue Wirick for support at the beamline and many helpful discussions. We thank Drs. Colleen Trantow and Sachiyo Iwashita for preparing melanosomes and Randy Nessler of the University of Iowa Central Microscopy Research Facility for technical assistance with freeze-drying of melanosome samples. The Central Microscopy Research Facility is supported by the Office of the Vice President for Research and Grant Number CA086862-10 to the Holden Comprehensive Cancer Center from the National Institutes of Health and National Cancer Institute. Biological aspects of this project were supported by National Institutes of Health grants to MGA, including Grant Number AR055697 from the National Institute of Arthritis and Musculoskeletal and Skin Diseases and EY017673 from the National Eye Institute. STXM imaging and associated travel were supported by University of Maine, the University of Heidelberg and the German Federal Ministry of Education and Research (BMBF) (project number 05KS7VH1) to MG and TH.
- Jacobsen C, Wirick S, Flynn G, Zimba C: Soft X-ray spectroscopy from image sequences with sub-100 nm spatial resolution. J Microsc. 2000, 197: 173-184. 10.1046/j.1365-2818.2000.00640.x.View ArticleGoogle Scholar
- Kirz J, Jacobsen C, Howells M: Soft X-ray microscopes and their biological applications. Q Rev Biophys. 1995, 28 (1): 33-103. 10.1017/S0033583500003139.View ArticleGoogle Scholar
- Numerical Python. [http://www.numpy.org]
- Matplotlib. [http://matplotlib.sourceforge.net/]
- IPython. [http://ipython.scipy.org/]
- Smith SW: The Scientist & Engineer's Guide to Digital Signal Processing. 1997, California Technical Publishing: San Diego, CAGoogle Scholar
- Lerotic M, Jacobsen C, Schäfer T, Vogt S: Cluster analysis of soft X-ray spectromicroscopy data. Ultramicroscopy. 2004, 100 (1-2): 35-57. 10.1016/j.ultramic.2004.01.008.View ArticleGoogle Scholar
- Press WH, Flannery BP, Teukolsky SA, Vetterling WT: Numerical Recipies in C++ (Second Edition). 2002, New York: Cambridge University PressGoogle Scholar
- Otsu N: A Threshold Selection Method from Gray-Level Histograms. IEEE Trans Sys Man Cyber. 1979, 9 (1): 62-66. 10.1109/TSMC.1979.4310076.View ArticleGoogle Scholar
- Axis 2000 software. [http://unicorn.mcmaster.ca/aXis2000.html]
- Petterson LGM, Agren H, Schürmann BL, Lippitz A, Unger WES: Assembly and Decomposition of Building Blocks to Analyze Polymer NEXAFS Spectra. Int J Quant Chem. 1997, 63 (3): 749-765. 10.1002/(SICI)1097-461X(1997)63:3<749::AID-QUA16>3.0.CO;2-#.View ArticleGoogle Scholar
- Benzerara K, Yoon TH, Tyliszczak T, Constantz B, Spormann AM, Brown GE: Scanning transmission X-ray microscopy study of microbial calcification. Geobiology. 2004, 2: 249-259. 10.1111/j.1472-4677.2004.00039.x.View ArticleGoogle Scholar
- Hitchcock AP, Morin C, Heng YM, Cornelius RM, Brash JL: Towards practical soft X-ray spectromicroscopy of biomaterials. J Biomater Sci Polym Ed. 2002, 13 (8): 919-937. 10.1163/156856202320401960.View ArticleGoogle Scholar
- Anderson MG, Haraszti T, Petersen GE, Wirick S, Jacobsen C, John SWM, Grunze M: Scanning transmission X-ray microscopic analysis of purified melanosomes of the mouse iris. Micron. 2006, 37 (8): 689-698. 10.1016/j.micron.2006.03.008.View ArticleGoogle Scholar
- Winn B, Ade H, Buckley C, Feser M, Howells M, Hulbert S, Jacobsen C, Kaznacheyev K, Kirz J, Osanna A, Maser J, McNulty I, Miao J, Oversluizen T, Spector S, Sullivan B, Wang Y, Wirick S, Zhang H: Illumination for coherent soft X-ray applications: the new X1A beamline at the NSLS. J Synchrotron Radiat. 2000, 7 (Pt 6): 395-404. 10.1107/S0909049500012942.View ArticleGoogle Scholar