A PC-based program for evaluation of Comparative Genomic Hybridization (CGH) experiments
互联网
Comparative genomic hybridization (CGH) is a molecular cytogenetic method for the detection of chromosomal imbalances. In a CGH experiment , two genomic DNA samples are simultaneously hybridized in situ to normal human metaphase spreads, and detected with different fluorochromes. The intensity ratio of the two fluorescence signals gives a measure for the copy number ratio between the two genomic DNA samples. For the objective identification of such imbalances, quantitative fluorescence digital image analysis is necessary. We present a complete system for CGH cytometry. It is based on a semi-automatic karyotyping system and runs under MS-Windows on an IBM-PC.
Keywords : Comparative genomic hybridization, CGH, image analysis, tumor cytogenetics, FISH
Introduction
Comparative genomic hybridization allows a comprehensive analysis of multiple DNA gains and losses in entire genomes within a single experiment. Genomic DNA from the tissue to be investigated, such as fresh or paraffin-embedded tumor tissue, and normal reference DNA are differentially labeled and simultaneously hybridized in situ to normal metaphase chromosomes. The DNA to be tested is labeled with biotin and is called test DNA . Genomic DNA derived from cells with a normal karyotyp is labeled with digoxigenin and serves as an internal control (control DNA). For the detection of the hybridized test DNA we used fluoresceine isothiocyanate and for the detection of control DNA anti-digoxigenin rhodamin. The fluorochromes were visualized by epifluorescence microscopy with selective filters. By comparing the fluorescence intensities of test and control DNA, changes in signal intensities caused by imbalances of the test DNA can be identified. The use of control DNA is necessary to have a local reference that can compensate signal variations in fluorescence intensities caused by some undesired reasons.
The basic assumption of a CGH experiment is that the ratio of the binding of test and control DNA is proportional to the ratio of the concentrations of sequences in the two samples.
To obtain quantitative, reliable, and reproducible results an accurate measurement of fluorescence intensities is necessary. Furthermore, many image operations must be performed to make digitized images of chromosomes in different metaphases comparable in order to improve the statistical fidelity of the detected genetic alterations by averaging. Thus, a quantitative fluorescence image processing system connected with a highly sensitive CCD camera is essential for CGH.
In this paper, we describe our basic approach to perform a CGH experiment and we discuss some aspects of hard- and software and show preliminary results.
Materials and Methods
The hybridization was performed as previously described with minor modifications. Normal chromosome metaphase spreads were prepared by standard procedures with special care to minimize cytoplasm debris. In general, no protein digestion was applied.
Prior to the hybridization, chromosome denaturation was performed for 60 sec at 77°C in 70% Formamid/2xSSC. Test and normal DNA, 5 µg each, were labeled by a standard nick translation reaction with Biotin and Digoxigenin, respectively. The DNase concentration was adjusted that the length of the labeled fragment varied between 200 and 1000 bp after 2 h incubation. One µg of each labeled genomic DNA, 20 µg of human Cot-1 DNA (Gibco BRL Life Technologies, Gaithersburg, MD, U.S.A) and 10 µg of herring sperm DNA were ethanol-precipitated.
The pellet was resuspended in 15 µl hybridization solution containing 33% Formamid/13.3% Dextransulfat/ 3xSSC, denatured at 77°C for 5 minutes and prehybridized at 37°C for 2 hours.
Finally, 12 µl of the hybridization mixture was applied to the slide with the denatured and dehydrated metaphase spreads. The hybridization was performed under a coverslip (18x18 mm) that was sealed with rubber cement for 3 days at 37°C. Detection of the differentially labeled genomic DNA was performed with avidin-FITC (Vector Laboratories, Burlingame, CA, U.S.A.) and anti-Digoxigenin-rhodamin (Böhringer Mannheim, Mannheim, Germany) followed by the DAPI staining for chromosome visualization. No enhancement reactions to increase the fluorescence signals of the genomic DNA probes were applied.
Images were acquired through a Zeiss Axiophot fluorescence microscope using a Plan NEOFLUAR oil objective x63, N.A. 1.25 (Zeiss, Oberkochen, Germany) equipped with filter sets appropriate for DAPI (Zeiss filter set 02, excitation: G365, beamsplitter: FT 395, emission: LP 420), FITC (Zeiss filter set 10, excitation: BP 450-490, beamsplitter: FT 510, emission: BP 515-565) and TRITC (Chroma filter set HQ Cy3+excitation filter from Zeiss filter set 15, excitation: BP 546/12, beamsplitter: FT 565, emission: BP 570-650) with a cooled CCD camera (Photometrics, Tucson, Arizona, U.S.A.) connected to a Macintosh Quadra 950 (U.S.A.). The resolution of this apparatus configuration is 0.108 m/pixel. The maximum image size is 1320x1035x12 bit.
The 100 W mercury lamp and the diaphragms of the microscope were precisely adjusted to get a homogeneous illumination of the optical field.
For each metaphase spread three gray-level images were digitized, one image for each fluorochrome. An image size of 512x512 or 768x768 pixels was chosen, according to catch the whole metaphase with one image. The images were inverted in order to make it possible to use the standard segmentation process and transferred as 8 bit TIFF-files to a server PC via a local area network.
Iver.Petersen@charite .de or Karsten.Schluens@charite.de for further information.
Running on a batch of metaphase images the program comprises the following steps: 1) image segmentation using the DAPI image; 2) correction of optical shift; 3) computation of the fluorescence ratio images between FITC and TRITC images; 4) karyotyping; 5) determination of chromosome axis; 6) stretching of chromosomes and calculation of profiles; 7) presentation of results
Image segmentation is a key step in image processing. It refers to the decomposition of an image into it's components, in our case into objects (chromosomes, interphase cells, etc.) and the background. Because of the good uniformity of the background it was possible to perform the segmentation by setting a single global gray value threshold THRS for the entire DAPI image. All pixels with a gray value above THRS were assigned to the object class and all other pixels to the background class. The best segmentation threshold THRS was determined interactively by changing the segmentation threshold and inspecting the segmented image which was visualized in real-time. After the best value for THRS was found a contour-following algorithm traced the boundaries of the objects resulting in closed contours for each object. The contours of singles objects can be corrected by the definition of individual thresholds by a contour correction dialogue. The final contours were used as segmentation masks for the FITC and TRITC images in the following steps.
A |
B |
D |
C |
Figure 1 Four images from the same metaphase of a small cell lung tumor. (A) DAPI image. (B) FITC image. (C) TRITC image. (D) Ratio image FITC/TRITC. The contours of the chromosomes were detected in the DAPI image and used as segmentation masks for the FITC and TRITC image.
It should be noted here that no classification of segmented objects was performed in this step. This means that touching chromosomes are clustered into one object and interphase cells and preparation artifacts are also regarded as objects.
To collect the three fluorescence images for each metaphase a mechanical movement of the filter slide was necessary. Mechanical imperfections by moving the filter slide manually caused local shifts between the three images. Because a CGH experiment is based on a measurement of intensity ratios between two fluorochromes, an accurate alignment of the three images is absolutely necessary. We chose to correct these image shifts interactively with software in the following manner.
The object contours found in the DAPI image were displayed as a colored overlay for the FITC image. Using the cursor keys of the PC keyboard the FITC image was moved until the best fit to the DAPI contours was reached. The same procedure was done with the TRITC image. During this correction process the user can switch very fast between DAPI, FITC, and TRITC image views using a Compare function. This allows for a highly sensitive control of image shifts, which can be corrected with the keyboard keys.
A |
B |
Figure 2 Correction of local shift. (A) FITC image with contours found in the DAPI image. (B) Corrected image.
After correcting the optical shift, the DAPI contours are also used for segmentation of the FITC and TRITC images.
Calculation of the FITC/TRITC ratio
For the detection of DNA-losses and -amplifications on the basis of the FITC/TRITC ratio it is necessary to normalize the average image brightness of the FITC and TRITC images. The FITC image gray values of all chromosome pixels (i.e. of all pixels within the segmentation masks) are inserted into a histogram. The statistical median value Cf is determined for this histogram. The FITC image gray values of all background pixels (i.e. of all pixels outside of the segmentation masks) are inserted into a second histogram. The statistical median value Bf is determined for this histogram. The background corrected median FITC intensity Mf is calculated as:
The same is done in the TRITC image for the calculation of the background corrected median TRITC intensity Mt . The background corrected gray value functions f(x,y) and t(x,y) are calculated as
where f '(x,y) and t '(x,y) are the intensities of the captured FITC and TRITC images at the position x,y. For all chromosome pixels a FITC/TRITC ratio image r(x,y) is calculated as:
r(x,y) := ( f(x,y) * 128 * Mt ) / ( t(x,y) * Mf ) for f(x,y) >= t(x,y)
r(x,y) := 256 - (( t(x,y) * 128 * Mf ) / ( f(x,y) * Mt )) for f(x,y) < t(x,y)
where f(x,y) is the background corrected FITC gray value and t(x,y) is the background corrected TRITC gray value at the image position x,y. The ratio values r(x,y) are clipped to the limits 0 and 255 resulting in a range of 0<=r(x,y)<=255.
Figure 3 depicts the color look-up table which indicates normal ratio values blue, DNA-amplifications green and DNA-losses red.
 |
Figure 3 Look-up table for pseudo coloration of FITC/TRITC ratio images |
A rough discrimination between objects and background was made in the image segmentation step. The next step was to discriminate the objects found in the DAPI image. This karyotyping process is controlled by the user, and necessary interactions are done manually by means of the mouse. The system karyotypes metaphase spreads with a high degree of automation. It is possible to work with the chromosomes in the monitor in the same manner as with a photography. The chromosomes can be cut out, cut up, glued together, moved and set at any angle. Touching chromosomes can be separated and artifacts and interphase cells can be deleted. Overlapping chromosomes and chromosomes influenced by artifacts were rejected. All operations which are done in the DAPI image, were simultaneous executed for the FITC and TRITC image automatically.
After finishing this segmentation step a karyogram form is drawn and the chromosomes are automatically vertically arranged. The chromosome classification, i.e. the position of the chromosomes in the karyogram form, can be carried out interactive or automatic with ensuing possibilities for interactive correction. The Compare function allows for switching between the DAPI, FITC, TRITC and FITC/TRITC ratio images. Especially the FITC/TRITC ratio image is a helpful tool for chromosome classification because the pseudo coloration of the ratio image makes it easy to identify homologous chromosomes.
A |
B |
D |
C |
Figure 4 Four CGH karygrams from the metaphase in Figure 1 . (A) DAPI image. (B) FITC image. (C) TRITC image. (D) Ratio image FITC/TRITC.
Determination of chromosome axis
The exact definition of the chromosome axis of symmetry is the basis for stretching chromosomes and for extracting the intensity ratio profiles. The first step was to perform a closing operation for the chromosome masks (dilation followed by an erosion of the segmentation masks) to smooth the chromosome contours. Holes within the chromosome masks were filled, which results in topological simple connected segmentation masks. Then the Hilditch skeleton was calculated for each chromosome. 3,7
The final chromosome axis was obtained by extending the skeleton tips. In spite of the foregoing smoothing process the skeletons can have short secondary branches and the ends of the skeleton may be misled in a wrong direction. Small disturbances of the skeleton can be corrected automatically but sometimes it is a difficult problem to find the main branch of the skeleton. In this case the user can define the chromosome axis interactively or can exclude this chromosome from the evaluation.
Figure 5 (A) Original chromosome from DAPI image. (B) Contour after smoothing and skeleton. (C) Main branch of the skeleton. (D) Polygon approximation and extension of the tips. (E) Sample lines for stretching chromosomes (only every fifth line is drawn).
Stretching of chromosomes and calculation of profiles
For the calculation of the perpendiculars of the chromosome axis an analytical description of the chromosome axis must be achieved. According to Groen et al. 2 the employment of piecewise-linear (PWL) approximation is preferred over the use of polynomial approximation techniques, because a chromosome is usually 'cracked' rather then bent. The chromosome axis is approximated by a polygon.
An array of points a[n] lying in unit distance on the chromosome axis polygon was created. Each array element a[i] has three entries:
x = x-coordinate of the point i
y = y-coordinate of the point i
a = direction of the perpendicular of the polygon edge corresponding to the point i
Abrupt changes of the direction at the polygon vertices v are smoothed:
1) The angle ß between neighbored polygon edges is calculated.
2) The direction difference ß is distributed over k points a[v-k/2]...a[v]....a[v+k/2] lying in the neighborhood of the polygon vertice v. The parameter k is chosen in such a way, that the perpendiculars of neighbored axis points are not crossing each other inside of the chromosome, i.e. k is a dependent of the angle ß and of the distance D of the axis point v from the chromosome contour: k := ß / arcsin( cos2( ß /2) / D ).
3) The angles of the axis points are corrected with the angle part d ß := ß / k for all i with 0<= i <= k/2:
a[ (v-k/2) + i ].a := a[ (v-k/2) + i ].a + (i * d) and
a[ (v+k/2) - i ].a := a[ (v+k/2) - i ].a - (i * d).
The pixels of the stretched chromosome are sampled along these perpendiculars using bilinear interpolation. Profiles can be calculated by averaging the gray values in the pixel lines of the stretched chromosomes.
This operation can be done in the DAPI, FITC, TRITC and FITC/TRITC ratio image.
Figure 5
To improve the statistical fidelity of the detected genetic alterations it is necessary to calculate average CGH karyograms and average chromosome profiles over a collection of metaphases. Because the length and the width of homologous chromosomes in different metaphase spreads are different, they were normalized to a predefined size using bilinear interpolation. The relative length of the different chromosome classes was chosen according to the ISCN definition.4 The relative width of the different chromosome classes was estimated on the basis of a random sample of 10 metaphases.
Every chromosome was included in the averaging process two times. One time in its original form and one time mirrored at the chromosome axis. Pixels of an accumulated chromosome were suppressed if the number of the accumulated pixels Np didn't fulfill the following condition:
where Nc is the number of accumulated chromosomes.
For good quality metaphases approximately 5-10 chromosomes for each class should be sampled in this averaging process.
Results
Alignement errors occurred between the three fluorescence images due to the use of different filters. We quantified the errors and measured the reproducibility of the correction procedure. The interobserver and intraobserver reproducibility of the interactive correction of optical shift is depicted in Table.1. The standard deviation of the correction of optical shift was approximative 0.5 pixel (i.e. 0.05 m). The maximum deviation of a single correction value from the optimum correction value was 1.4 pixel (the corresponding total mean was assumed to be the optimum value).
| metaphase 1 | metaphase 2 | metaphase 3 | |
| dx dy | dx dy | dx dy | |
| observer 1 | -3.0 ± 0.0 0.5 ± 0.6 | -4.0 ± 0.0 -1.0 ± 0.0 | -3.8 ± 0.5 0.0 ± 0.8 |
| observer 2 | -3.0 ± 0.0 0.0 ± 0.0 | -3.8 ± 0.5 -1.0 ± 0.0 | -3.5 ± 0.6 0.3 ± 0.5 |
| observer 3 | -3.0 ± 0.0 0.5 ± 0.6 | -3.8 ± 0.5 -0.3 ± 0.5 | -3.5 ± 1.0 0.3 ± 0.5 |
| total | -3.0 ± 0.0 0.3 ± 0.5 | -3.8 ± 0.4 -0.8 ± 0.5 | -3.6 ± 0.7 0.0 ± 0.6 |
Table 1 Interobserver and intraobserver differences (in pixels) of interactive determined correction of
optical shift between the FITC and TRITC images of three different cases.
For the detection of DNA-losses and -amplifications on the basis of the FITC/TRITC-ratio it necessary to normalize the average image brightness of the FITC and TRITC images. It is possible to produce relatively homogenous FITC and TRITC images by optimum adjustments of the optical conditions. In the case of homogenous images, the described simple normalization method provides good results. Figure 7 depicts the influence of the correction of optical shift and the normalization of the mean image brightness on the intensity profile of the FITC and TRITC images of a typical metaphase. The high correlation between the FITC and TRITC background signals enables the applicability of the CGH method also for regions with inhomogenous background signals due to the presence of cytoplasm.
Figure 7 FITC and TRITC intensities on a sample line. (A) Without correction. (B) After shift correction (dx=-2, dy=-1). (C) After shift correction and brightness normalization (Bf=16, Cf=132, Mf=116, Bt=20, Ct=116, Mt=96).
Figure 8 shows the limits of the described method for brightness normalization. Incorrect illumination caused a decrease of the background signal of the FITC image from left to right and an increase of the background signal of the TRITC image in the same direction. It should be possible to correct this illumination error by fitting a plane to the background signals of the FITC and TRITC images, but generally such a complicated and error prone correction can be avoided, if the optical apparatus is carefully adjusted.
Figure 8 FITC and TRITC intensity profiles after shift correction and brightness normalization. (B) Corresponding FITC/TRITC ratio image. The images were captured with incorrect illumination. Note that the background signal of FITC on the left side is higher than for TRITC and vise versa on the right side.
Figure 12 depicts a typical result of a CGH experiment. It shows a karyogram with the normalized chromosomes in form of a ratio image. To each chromosome belongs the averaged 1-dimensional ratio profile. Figure 12 was obtained by analyzing 10 metaphases of one case from a patient with SCLC. The metaphase spreads were analyzed, and the karyogram of the averaged TRITC/FITC ratio chromosomes and the FR profiles of all metaphase spreads were drawn.
The three vertical lines at the right side of each chromosome indicate thresholds defining the three intervals of balanced, underrepresented, and overrepresented chromosomal material. Note that these thresholds were arbitrarily defined under the assumption that 50 % of the cells in a diploid tumor cell population carry a chromosomal imbalance. Then for a monosomy the theoretical ratios are 0.75 and for a trisomy 1.25. The usefulness and robustness of these thresholds was shown in 11 . A brief discussion about the limitations of this attempt and a suggestion for other thresholds are made in 12 .
The profiles are extracted by adding the ratio values of the pixels located in lines perpendicular to the chromosomal axis inside of the chromosome divided by the number of these pixels. The central line represents the balanced state. If the mean profile at a locus is higher than the right threshold, there is an amplification of DNA. If it is less than the left threshold, there is a loss of chromosomal DNA.
Together with the profiles the fluorescence ratio images of each chromosome give a good survey about DNA gains and losses in the tumor. Using the color-look-up-table in Figure 3 chromosomes are displayed in pseudo colors showing the strength of genetic imbalances. An area colored light-green represents high ratio values corresponding to amplifications. Whereas an area which is colored dark green represents a low-level amplification. Losses are shown as shades of red and blue color corresponds to the balanced state of the chromosome material. Especially for small areas of amplifications or losses, the visual inspection of the pseudo colored pictures of chromosomes improves the evaluation and interpretation of a CGH experiment.
Figure 4 ). (A) ,(B) homologous chromosomes. (C) Mean chromosome and mean profile from chromosomes (A) and (B). (D) chromosome three of another metaphase in the same case, without width normalization.
In general we used FITC for visualizing the test DNA and TRITC for visualizing the control DNA. In order to test the errors due to possible differences in the preparation process we investigated the reversed painting scheme. Figure 10 shows the mean profile of chromosome 3 averaged over 8 metaphases of two CGH experiments with reversing painting schemes using the same tumor DNA. The ratio profiles were calculated individually for each chromosome and then averaged, resulting in a mean profile. To test the fidelity of results we estimated the 95% confidence interval. The standard error was calculated for each pixel and chromosome along the profiles. We calculated the 95% confidence interval by adding the standard error two times to the mean profile and subtracting the standard error two times from the mean profile.
As shown in Figure 11 , the test indicated little differences between both experiments. The differences between the detected amplification and losses lie inside the confidence interval. This gives us the certainty that our hybridization regime permits quantitative interpretation of ratio changes.
Figure 10 Mean chromosomes and mean FITC/TRITC profile of chromosome three from eight metaphases with changed painting scheme. (A) Test DNA visualized by FITC. (B) Test DNA visualized by TRITC. (The TRITC/FITC ratio image and ratio profile is depicted instead of the FITC/TRITC ratio to ease the comparison with other figures)
A |
B |
Figure 11 mean profile with 95% confidence interval. A test DNA wih FITC, B test DNA wih TRITC
Figure 12 shows the final result of a CGH experiment. It represents the analysis of 10 metaphases of a small cell lung carcinoma in the form of a ´CGH karyogram´. Typical aberration of this tumor entity include loss of chromosome arm 3p with simultaneous overrepresentation of 3q and DNA losses on chromosome 5q, 13q, and 17p (Ried, Petersen et al. 1994). In this tumor regions of underrepresentation are additionally seen for chromosomes 2p, 4q, 10, 15, 16q and 22 whereas overrepresentations occured for chromosomes 1, 2q, 7, 8, 12 14 and 18. This multitude of changes correlates to the fact that small cell lung carcinoma is one of the most malignant tumors in man.
Figure 4
With a CGH experiment chromosomal imbalances in entire genomes can be detected quantitatively. To do this requires dealing with a lot of complex issues. The program described in this paper is based on an existing program for karyotyping. For the evaluation steps it is advantageous if the chromosomes were organized in a karyotyp.
Speed, easy usability and accuracy are the main performance criteria for the CGH program. The generation of a CGH karyogram (including correction of optical shift, segmentation and classification) by an experienced user takes about 10 to 15 minutes.
The segmentation of the chromosomes is performed by manually setting a single global gray value threshold for the entire DAPI image. The interactive thresholding method is preferred to an automatic histogram based thresholding method because its superiority has been proved in many investigations.8,9 The limitations of histogram methods are usually connected with the fact that the histogram structure and image contents have little in common (since histograms do not include any information about local relations in the image). It should be investigated if segmentation methods which include global information and apriori knowledge can be used for automatic chromosome segmentation.1,5,10
The representation of the chromosomes in form of a CGH karyogram facilitates the evaluation of the CGH experiment. In the evaluation process of a CGH experiment the user can add as many metaphases he or she wants. Averaging the pictures and profiles from several homologous chromosomes improves the statistical fidelity of the detected genetic alterations. This averaging will stop systematic mistakes that could lead to a wrong interpretation of the results. It leads to smooth local variations caused by random noise and reduces the influence of errors that were caused from incorrect lighting or mistakes in the recording or in the analysis of the image. However, it might decrease the resolution capability of a CGH experiment. The resolution of DNA gains and losses in the test genome is highly influenced by the quality of the fluorescence images and the algorithms used for chromosome averaging.
The result in form of a mean CGH karyogram together with the averaged ratio profiles is shown at all times. The view of the CGH karyogram can be switched between DAPI, FITC, TRITC and FITC/TRITC ratio.
At each step the program allows the user to visually inspect all chromosomes which are included in the averaging process. Chromosomes can be then excluded from the evaluation if they are incorrectly segmented, are incorrectly stretched or are unsuitable for the evaluation due to other causes.
The program allows not only for the calculation of mean chromosomes and the standard deviation, but also for statistical tests to be carried out in order to test the significance of an alteration. Through statistical tests it is possible to detect obvious outliers, that, for example, indicate faulty digital image processing or misclassification, and exclude them from the calculation.
As mentioned above about ten good quality metaphases should be evaluated to get stable results for a CGH experiment. Nevertheless, the optimal number of metaphases that are needed to be evaluated for reliable CGH analysis depends on the variability of the data. The higher the variability the more metaphases must be evaluated. This is true only to a certain extent, from where a CGH analysis makes no more sense. The variability critically depends on the quality of the hybridization and the digital images and the image processing . In our opinion it should be possible to use a statistical test based on the measured data to determine the number of metaphases needed to be evaluated for an actual case.
Up to now only a visual inspection of the results of a CGH experiment is possible using the final ratio image and the ratio profiles. We are working on an automation of data interpretation that is statistically validated. This is desirable for reasons of reproducibility that is the basis for comparing results. Especially for interpreting a so called super karyogram, a statistically validated interpretation is necessary.
The overall result of a CGH experiment depends on the performance of each step. Every step in the preparation and the image analysis must be optimized in order to obtain reproducible results. With a greater exactness, the resolution capability of the CGH is raised. That means that it would be possible for changes with lower local extensions and lower strengths to be detected.
In the same way that it is possible for a case to be ascertained over several metaphases, it is also possible for a super karyogram to be generated through the combination of several cases of a tumor entity.(summary of a series of CGH experiments of a particular tumor entity).
Through the comparison of individual cases with super karyograms of different tumor classes it should be possible to carry out a classification of the individual case. It is important here to once again emphasize that this can only function if a successful complete standardization of all the steps of a CGH experiment can be realized including an automated statistically validated interpretation of the results.
