| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
From the Laboratory for Cytochemistry and Cytometry,
*
Department of Molecular Cell Biology, the Department of Human Genetics,
and the Department of Pathology,
Leiden University Medical Center, Leiden, The Netherlands
 |
Abstract |
|---|
 Top Abstract IntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
|
Introduction |
|---|
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
Compared to conventional visualization, the speed of molecular analysis is increased by a factor of more than a hundred. Construction of TMAs, originally done by hand, can be automated, increasing the throughput even more. At present, limiting factors in the use of TMAs for high-throughput analysis relate to the visual evaluation of the staining patterns of each of the individual cores in the array and to the subsequent input of the results into a database. Such a database is essential to correlate the data with tumor type and outcome, and to evaluate the performance against other markers, achieved in separate experiments. So far, these steps are mostly performed by hand, and are consequently time-consuming and potentially prone to bias and errors, respectively.
This problem has been recognized and, recently, attempts have been made to use automation and image analysis to address this issue.18, 19 Our approach in addressing these issues is to use digital scanning techniques and subsequent low-resolution image analysis to generate a classification code that is automatically stored in databases. The system is based on a flat-bed-type scanner ("top of the line color photocopier") that records images at zero optical magnification with a pixel resolution of about 5 x 5 µm2. It provides digital zooming of arrays for interactive verification of the results and rapid linkage to data sets of other markers derived from the same array. Performance of the system was compared to manual classification for a representative set of arrays containing colorectal tumors stained with different markers.
|
Materials and Methods |
|---|
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
Furthermore, the arrays were analyzed for ki-67 (clone MIB1; DAKOCytomation, Glostrup, Denmark), P53 (clone DO-7; DAKOCytomation), ß-catenin (clone 14; Transduction Labs, Lexington, KY, USA) and cytokeratin (Pan Ab-1, clone AE1/AE3; NeoMarkers, Fremont, CA, USA) staining. Cytokeratin analysis was included to automatically omit punches without cytokeratin positivity (epithelial fraction not present), thus preventing false-negative staining results.
P53 staining was manually classified according to the following four categories: category 1 when no cells showed nuclear staining (possibly mutated p53 or due to eg, mdm2 overexpression), category 2 for 0 to 25% staining (predicted wild-type p53), category 3 for 25 to 75% staining (possibly mutated p53), and category 4 for more than 75% of the cells showing strong nuclear staining (predictive for mutated p53). Ki67 staining was classified in a similar way. The four categories of ß-catenin staining were distinguished as follows: category 1 when only membranous staining was seen, category 2 for nuclear staining on the peripheral edges of tumor, category 3 for elevated nuclear staining and membranous staining, and category 4 for strong nuclear expression and weak membranous staining.
Hardware Configuration
The system consists of a flat-bed Agfa XY-15 scanner interfaced to a 933MHz Power Mac G4 computer via a SCSI-2 interface. The computer is equipped with 512 MB memory and a 60 GB disk for image storage. The single pass XY-15 scanner can be used both in transparency and reflective mode, from which the former is applied for digitizing the microscopic slides. It has a tri-linear 8000 element charge coupled device (CCD) provided with optical XY-zoom technology. In this way an optical resolution of 5000 dots per inch (dpi) is obtained in both directions corresponding to about 5.0 µm. Sampling density can be increased further to 15000 dpi by means of interpolation. The maximum scan area is 350 x 455 mm, corresponding to A3+ format. However, the highest optical resolution is only achieved within the middle part of 232 x 455 mm. A special mold, which can accommodate 45 microscopic slides, has been constructed to fit on this part of the scanner. All slide holders are equipped with springs that press a slide against two notches, so that each slide position is well defined. The maximum density range of the scanner is 0 to 3.9, resulting in a color depth of 16 bit per Red Green Blue (RGB) color. Images of tissue microarrays and individual cores, including their derived features, can be archived on a large file server for reviewing the data via a local area network.
Image Acquisition
Stained sections were digitized using the ColorExact software package from Agfa, which additionally allows the setup of the scanner characteristics, such as scan mode, scan resolution, color calibration, and dynamic range. Usually a bed-scan of the complete scan area is performed first at a relatively coarse resolution of 100 dpi. Then the bed-scan is automatically analyzed for the presence of microscopic slides and for each slide the label and specimen part are determined. This provides the necessary information about the orientation of the tissue microarrays and the coordinates of the regions for scanning the arrays at a higher resolution. A batch job is thereby generated for the automatic scanning of these regions by ColorExact, so that up to 45 slides can be digitized automatically.
The tissue microarrays for this study were all sampled at about 5600 dpi just above the theoretical spatial resolution. Typical image sizes of 50 MB were obtained for tissue microarrays of 24 x 16 cores, while the images were scanned at a speed of about 30 MB per minute. Additional time is necessary to calibrate the scanner and to focus an image crop and takes 30 to 40 seconds. Autofocusing is based on an interpolation scheme, which is an integral part of the ColorExact software, and cannot be influenced by the operator. This focus procedure appeared to be sufficient for this study.
Image Analysis
The analysis of the TMAs was performed automatically and can be divided into the following steps: the determination of the grid size, the localization of the individual cores, the measurement of image features, and the evaluation and archiving of individual core data.
For segmenting the individual cores the color TMA image was converted to a gray-value image by equally averaging the separate RGB channels. On the basis of the gray-value histogram, an initial global threshold was determined to segment the individual cores from the background. The binary image was cleaned by removing small particles and partially detected cores utilizing morphology operations such as erosions and openings. The remaining objects were assumed to be tissue cores and features, such as size, center of gravity, and shape (eccentricity) were measured, while for each object the distances to its four nearest neighbors and the corresponding angles also were determined. The core size of the grid was calculated as an average from the objects, which were sufficiently round and was used later as an estimate for the detection of cores, which were not found initially or were incomplete.
A histogram based on the distances of the individual cores to their nearest neighbors was calculated to estimate the intercore spacing of the grid. The tissue microarrays analyzed in our laboratory showed little difference in spacing along the two main axes and did not result in significantly different peaks. The first peak in this histogram, in most cases also the highest peak, corresponds to distances belonging to 4-connective neighboring cores. For the distances, which lay within twice the SD of this peak, the angular histogram was also determined and corrected for eventual shifts of 180 degrees. The two highest peaks in this latter histogram, mostly at a distance of approximately 90 degrees from each other, correspond to the directions of the main axes. More accurate values of these directions were calculated by averaging only the angles, which lay within the SD of a peak value. Finally, the precise grid spacing in both directions separately was determined by averaging the distances belonging to the two peaks within the angular histogram separately.
This method for the determination of the grid characteristics requires that a sufficient number of cores are initially detected, so that mainly distances of those nearest neighbors are found, that are 4-connective. For the tissue microarrays, analyzed in our laboratory, this did not seem to present a problem.
The grid characteristics found in combination with the detected cores of the correct shape and size were used to fit an initial grid for the TMA. The positions of the remaining cores were determined by local bilinear interpolation and, when necessary, by predictive extrapolation. When the position of an interpolated core almost coincided with the position of an object that was initially detected but was not of the correct size and/or shape, the core position was adjusted to optimally cover the object area.
When cores are not exactly matched automatically, the operator still has the possibility to adapt the core positions interactively, so that the correct features are measured. In the TMA slides analyzed so far no real subarrays were present. Partially empty rows showed the same spacing as normal rows, so that no mismatches occurred as empty cores were fit in between. In principle, the program may be adapted so that also larger and more complex TMA slides may be handled.
Feature Extraction and Image Archiving
Since the resolution of the flat-bed scanner is not sufficient enough to analyze the individual cells within the tissue cores, only some global features were extracted to characterize each core. The cores were thereby segmented into parts containing probe signal and counterstain, while also empty or partially empty cores were found. In addition to area percentages, the total absorption of both the probe and counterstain parts were determined by summing the logarithmic intensity values corrected for the logarithmic value of the averaged background intensity of the corresponding pixels. Other parameters were also measured, such as peak intensity and parameters derived from the intensity histogram, to characterize the distribution of the intensities within the probe and counterstain class.
The numerical data were saved in an Excel spreadsheet to classify the different tissues and were organized as follows. A directory structure is created for each type of tissue array. The main directory contains a spreadsheet for each measured marker and a separate spreadsheet describing the organization and the content of the array. A subdirectory for each individual core is created where the corresponding images for the various markers reside. The measurements of the individual cores are organized row-wise in the spreadsheet, while a column is reserved for the file path to the corresponding image. Spreadsheets as generated by the pathologists containing the visual classification of the arrays in the traditional four classes can be easily combined with the spreadsheets of the measured data using a simple macro. In principle, an additional macro also can be written to reformat the data, so that it can be applied to hierarchical cluster analysis programs for tissue microarrays, as described recently by Liu et al.18
The analysis program itself can also retrieve both the measurement and image data for the different markers so that corresponding cores from different markers can be displayed simultaneously. This is considered very useful because it provides the pathologist with direct visual information as to how the expression of a set of markers differs for various cancer types.
|
Results |
|---|
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
the various steps of the image analysis procedure are shown to determine the grid and the position of the individual tissue cores on a microarray stained for p53. In all microarrays analyzed up to now, it appeared that after morphological filtering to reduce the noise particles (Figure 1b)
, enough tissue cores remained detectable to determine reliable estimates for the direction of the axes of the grid and for the spatial distance between neighboring cores. The interpolation algorithm for fitting the remaining cores, which were not round enough or were incomplete, also appeared to function satisfactorily. Only in rare cases, when a very small proportion (less than 10%) of a core was present and when it was lying considerably off-axis, human interaction seemed to be necessary to correct the automatically determined position.
|
As the manual classification in four classes is based on the intensity of the probe signal and the proportion of the tissue that shows expression, the integrated absorption and the relative area of the probe signal were used to classify the tissue cores. In Figure 2
, the results are shown for the nuclear markers, p53 and ki67, and the membrane marker ß-catenin. The measurements on the nuclear markers show a good correlation with the visual classification in four classes by the pathologist, although the classes largely overlap. Correlation coefficients of 0.75 and 0.77 were measured for, respectively, the area and the integrated absorption of marker p53 and of 0.66 and 0.72 for ki67. Such a correlation is not present for the membrane marker ß-catenin, as correlation coefficients of 0.10 and 0.21, respectively, were obtained. In Figure 2d
, the absorption distribution is shown as a function of the total absorption for p53 expression. The absorption distribution is a parameter derived from the absorption histogram by non-linear weighting of the various absorption classes. However, this parameter or other parameters which were measured, such as peak intensity, average intensity of the marker, and relative marker area, did not significantly improve the classification results.
|
, corresponding cores of tissue microarray stained with p53, ki67, and ß-catenin are shown illustrating the resolution that can be achieved by flat-bed scanning. In Figure 4
, part of a tissue core stained for p53 is electronically zoomed out to the cellular level and compared to the corresponding microscopic image digitized using a 10x objective. This clearly illustrates the limited resolution of the flat-bed scanner. Cells appear as out-of-focus blobs and cannot be individually recognized.
|
|
|
Discussion |
|---|
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
Despite the rather low resolution of the flat-bed some global measurements are performed to classify the individual cores. The results of this pilot study are encouraging. Especially for nuclear markers, such as p53 and ki67, requiring relatively low spatial resolution, classification based on the integrated absorption and the probe area show a similar trend when compared to the visual classification by the pathologist, although the classes still show a large overlap. For membrane markers, such as ß-catenin, which requires higher spatial resolution, the resolution of the flat-bed scanner appears to be too low to provide similar results. Some improvements were obtained by combining the measurements of the markers with the presence of sufficient amount of cytokeratin positivity in the TMA of a neighboring cross-section. This eliminates cores for which the integrated absorption is based on a relatively small proportion of the cancerous tissue. A recent and more extensive study by Camp et al19 has shown that automatic analysis of tissue microarrays has a better reproducibility than visual pathologist-based scoring, although the continuous scale of automated classification tends to overlap with the visual-based scoring in four discrete classes. However, it shows that for ß-catenin expression in TMAs of colon cancer, automated analysis of subcellular compartments is able to discern subtle differences in staining intensity resulting in new subsets of tumors with prognostic significance, which is not seen using visual pathologist-based assessment.
A separate study to evaluate the visual classification of TMAs by comparing the staining of whole tumor slides versus the results on TMAs showed that in the case of p53 and ki67 the four different classes could be reliably recognized in TMAs. However, the visual classification for ß-catenin-stained TMAs appeared also to be much more difficult; since only category 4 corresponding to a strong nuclear expression and weak membranous staining could be reliably confirmed in whole tumor slides, while the three other categories could not easily be distinguished in the TMAs as separate staining patterns.
Currently the tissue cores are segmented into areas containing the expression marker, the counterstain, and eventually in areas with no cellular material present and absorption measurements are performed on these different parts. It was hypothesized that a better method might be to calculate the absorption through calibrating the exact colors of the counterstain and the expression marker using reference cores, and then integrate the contributions per pixel to the different classes. However, this method did not result in a significantly better classification. We also tested whether over-sampling above the optical resolution to 15000 dpi would improve the results. Indeed, the over-sampled images of the individual tissue cores looked better on the screen, when compared to those digitized at the optical resolution and re-sampled at the same interpolated resolution using the bi-cubic interpolation scheme of Photoshop. However, the classification results of the TMAs showed no improvement. Over-sampling has almost no influence on the acquisition time and increases the analysis time only marginally. However, the database will grow by a factor of 6 due to the increased image size of the cores.
The analysis program is capable of automatically processing the TMAs without any human interaction. Only tissue cores, which are incomplete, are sometimes not optimally covered by the bilinear interpolation and predictive extrapolation algorithm. An algorithm, to optimize the fitting between cores, as found by interpolation and the amount of tissue present in the direct vicinity, could probably improve this further, so that no human interaction would be necessary even when the initial fit is not optimal.
It appeared to be very useful to combine the analysis of TMA slides with hands-off data input in a database for visualization of the tissue cores. It provides the pathologist with direct visual feedback on how the expression signals can differ for different types of cancer and it shows also how the expression patterns may vary for patients with the same type of tumor. Furthermore, it avoids errors related to the manual input of data that are considered substantial. Realizing that the use of TMA easily leads to input of thousands of data points, automated classification and data input seem to be the only solution.
The resolution of the Agfa scanner appeared to be sufficient for the analysis of the nuclear markers, while it was still too low to classify the tissues stained with membrane markers and to identify the individual cells visually. This would require at least a resolution improvement of a factor of 3 to 4. However, it is unlikely that the resolution of flat-bed scanners will improve in the near future. The optical resolution of the Agfa XY-15 is already among the highest currently available and is below the grain size as used in photographic films, so that the need for higher resolution is hardly present for the graphic industry, where most of these systems are sold.
On the other hand, a similar approach for the processing of TMAs could also be followed using a microscope system with an automated scanning stage, objective revolver, autofocus, CCD camera, and, optionally, a slide feeder, when large numbers of TMAs have to be analyzed. A TMA could first be scanned at a relative low magnification (using a 2.5 to 10x objective) to determine the grid characteristics and the positions where the individual cores are located. The individual tissue cores could then be acquired at a higher magnification (using a 20 to 40x objective) by stitching separate CCD images together when necessary. At present, commercial systems including automated slide loading are available. The algorithms for automated processing of TMA slides at low magnification as described in this paper and image acquisition of individual cores at high magnification can easily be implemented in such systems. The program in its current form is, however, available for the scientific community.
The automated microscopic approach for analyzing tissue microarrays will result in much larger images for the tissue cores (about 1 to 5 MB at 0.5-µm resolution), but with the ongoing development of faster PCs and steadily increasing data storage facilities it is feasible to easily store millions of tissue sections into a database. This provides the pathologist not only with the ability to directly retrieve and visualize the expression patterns for various cancers, but also with the possibility to zoom into the cellular level.
|
Acknowledgments |
|---|
|
Footnotes |
|---|
Supported by the "Anna en Maurits de Kock Foundation."
Accepted for publication March 27, 2003.
|
References |
|---|
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
expression associates with high proliferation rate and poor prognosis in oligo-dendrogliomas. Neuropathol Appl Neurobiol 2000, 26:504-512[Medline]
This article has been cited by other articles:
 |
|
 |
|
  S. Cheng, W. Liu, M. Mercado, S. Ezzat, and S. L Asa Expression of the melanoma-associated antigen is associated with progression of human thyroid cancer Endocr. Relat. Cancer, June 1, 2009; 16(2): 455 - 466. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 L. H. SCHMIDT, S. BIESTERFELD, A. KUMMEL, A. FALDUM, M. SEBASTIAN, C. TAUBE, R. BUHL, and R. WIEWRODT Tissue Microarrays are Reliable Tools for the Clinicopathological Characterization of Lung Cancer Tissue Anticancer Res, January 1, 2009; 29(1): 201 - 209. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 R. J. Marinelli, K. Montgomery, C. L. Liu, N. H. Shah, W. Prapong, M. Nitzberg, Z. K. Zachariah, G. J. Sherlock, Y. Natkunam, R. B. West, et al. The Stanford Tissue Microarray Database Nucleic Acids Res., January 11, 2008; 36(suppl_1): D871 - D877. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 M. Uhlen, E. Bjorling, C. Agaton, C. A.-K. Szigyarto, B. Amini, E. Andersen, A.-C. Andersson, P. Angelidou, A. Asplund, C. Asplund, et al. A Human Protein Atlas for Normal and Cancer Tissues Based on Antibody Proteomics Mol. Cell. Proteomics, December 1, 2005; 4(12): 1920 - 1932. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
M. Uhlen and F. Ponten Antibody-based Proteomics for Human Tissue Profiling Mol. Cell. Proteomics, April 1, 2005; 4(4): 384 - 393. [Abstract] [Full Text] [PDF]
|
|
| |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
