0 1994 Wiley-Liss, Inc. Cytometry 16:7-16 (1994) Automatic Fluorescence MetaDhase Finder h e e d s Translocation Scoring in FISH 6ainted Chrom&omesl J. Piper, M. Poggensee, W. Hill, R. Jensen, L. Ji, I. Poole, M. Stark, and D. Sudar MRC Human Genetics Unit, Edinburgh, Scotland, (J.P., W.H., L.J., I.P., M.S.); Division of Molecular Cytometry, Department of Laboratory Medicine, University of California, San Francisco (M.P., R.J., D.S.), and Lawrence Berkeley Laboratory, Berkeley (J.P.), California Received for publication September 2, 1993; accepted November 18, 1993 A fluorescence metaphase finder was constructed with commercially available hardware and a standard Unix worksta- tion. Its accuracy was measured in terms of the number of false positive and false negative detected metaphases on a vari- ety of different slide preparations. The metaphase finder was used in a translo- cation scoring experiment in which meta- phase preparations of human peripheral blood lymphocytes were hybridized with whole chromosome probes to chromo- somes #1, #2, and #4. The automatic finder presented metaphases to the cyto- geneticist, centered in the eyepieces at x63. The cytogeneticist’s scores of ana- lyzable metaphases and of painted chro- mosomes involved in rearrangements were recorded. The time for the analysis was recorded and compared to the time to analyze a similar number of cells in a purely visual experiment in which the cy- togeneticist scanned for cells and ana- lyzed them, both at x63. The results showed that, neglecting the machine time spent scanning unattended, the amount of time required for the analysis was re- duced by a factor of three. Furthermore, in this experiment the metaphase finder found more scorable metaphases than the cytogeneticist found by visual scan- ning. Machine-assistedscoring had addi- tional, less quantifiable,benefits; notably that digital images of metaphases some- times assisted the analysis of chromo- some rearrangements, that cells could be revisited easily, and that the analysis was much less fatiguing. o 1994 Wiley-Liss, Inc. Key terms: Aberration scoring, CCD camera, chromosome, image analysis, in situ hybridization, region of interest analysis, scanning Automated metaphase finding and retrieval at high magnification for visual scoring has long been recog- nized as being of major benefit in the scoring of random chromosome aberrations (8). Finnon et al. (5) showed that the total cytogeneticist time for scoring a large population of metaphases for the “classical” radiation induced aberrations-dicentric and ring chromosomes, and acentric fragments-was reduced to one-third if a metaphase finder was employed. Shafer et al. (18) reached a similar conclusion in the context of scoring sister chromatid exchanges. With the advent of whole chromosome fluorescent in situ hybridization probes (“chromosome paints”), a rapid new assay has recently become available for de- tecting random chromosome rearrangements (4, 12, 13). Metaphases in which one or a few chromosome types have been painted are scored for disturbances to the normal complement of painted chromosomes; these typically present as the occurrence of bi-color chromo- somes composed of a painted and a counterstained part (Fig. 1). Except in the case of huge clastogenic expo- sure, by far the most frequently scored outcome is the absence of any such aberration in a metaphase. Since the “normal” pattern of these negative metaphase cells is easily recognized, visual scoring is extremely rapid. Apart from the advantage of speed, chromosome painting permits easy detection of the stable reciprocal translocations (Fig. 1). These are exceedingly difficult ‘Support for this work was provided by NCI Interagency agreement no. YO1-CP-10561 and by the Office of Health and Environmental Research, US. Department of Energy under contract DE-AC-03- 76SF00098. ‘Address reprint requests to Ron Jensen, Division of Molecular Cy- tometry, Department of Laboratory Medicine, MCB-230, University of California, San Francisco, Box 0808, San Francisco, CA 94143- 0808.
Transcript
0 1994 Wiley-Liss, Inc. Cytometry 16:7-16 (1994)
Automatic Fluorescence MetaDhase Finder heeds Translocation Scoring in FISH 6ainted Chrom&omesl
J. Piper, M. Poggensee, W. Hill, R. Jensen, L. Ji, I. Poole, M. Stark, and D. Sudar MRC Human Genetics Unit, Edinburgh, Scotland, (J.P., W.H., L.J., I.P., M.S.); Division of Molecular Cytometry,
Department of Laboratory Medicine, University of California, San Francisco (M.P., R.J., D.S.), and Lawrence Berkeley Laboratory, Berkeley (J.P.), California
Received for publication September 2, 1993; accepted November 18, 1993
A fluorescence metaphase finder was constructed with commercially available hardware and a standard Unix worksta- tion. Its accuracy was measured in terms of the number of false positive and false negative detected metaphases on a vari- ety of different slide preparations. The metaphase finder was used in a translo- cation scoring experiment in which meta- phase preparations of human peripheral blood lymphocytes were hybridized with whole chromosome probes to chromo- somes #1, #2, and #4. The automatic finder presented metaphases to the cyto- geneticist, centered in the eyepieces at x63. The cytogeneticist’s scores of ana- lyzable metaphases and of painted chro- mosomes involved in rearrangements were recorded. The time for the analysis was recorded and compared to the time to analyze a similar number of cells in a purely visual experiment in which the cy-
togeneticist scanned for cells and ana- lyzed them, both at x63. The results showed that, neglecting the machine time spent scanning unattended, the amount of time required for the analysis was re- duced by a factor of three. Furthermore, in this experiment the metaphase finder found more scorable metaphases than the cytogeneticist found by visual scan- ning. Machine-assisted scoring had addi- tional, less quantifiable, benefits; notably that digital images of metaphases some- times assisted the analysis of chromo- some rearrangements, that cells could be revisited easily, and that the analysis was much less fatiguing. o 1994 Wiley-Liss, Inc.
Key terms: Aberration scoring, CCD camera, chromosome, image analysis, in situ hybridization, region of interest analysis, scanning
Automated metaphase finding and retrieval at high magnification for visual scoring has long been recog- nized as being of major benefit in the scoring of random chromosome aberrations (8). Finnon et al. (5) showed that the total cytogeneticist time for scoring a large population of metaphases for the “classical” radiation induced aberrations-dicentric and ring chromosomes, and acentric fragments-was reduced to one-third if a metaphase finder was employed. Shafer et al. (18) reached a similar conclusion in the context of scoring sister chromatid exchanges.
With the advent of whole chromosome fluorescent in situ hybridization probes (“chromosome paints”), a rapid new assay has recently become available for de- tecting random chromosome rearrangements (4, 12, 13). Metaphases in which one or a few chromosome types have been painted are scored for disturbances to the normal complement of painted chromosomes; these
typically present as the occurrence of bi-color chromo- somes composed of a painted and a counterstained part (Fig. 1). Except in the case of huge clastogenic expo- sure, by far the most frequently scored outcome is the absence of any such aberration in a metaphase. Since the “normal” pattern of these negative metaphase cells is easily recognized, visual scoring is extremely rapid.
Apart from the advantage of speed, chromosome painting permits easy detection of the stable reciprocal translocations (Fig. 1). These are exceedingly difficult
‘Support for this work was provided by NCI Interagency agreement no. YO1-CP-10561 and by the Office of Health and Environmental Research, US. Department of Energy under contract DE-AC-03- 76SF00098.
‘Address reprint requests to Ron Jensen, Division of Molecular Cy- tometry, Department of Laboratory Medicine, MCB-230, University of California, San Francisco, Box 0808, San Francisco, CA 94143- 0808.
8 PIPER ET AL.
FIG. 1. A metaphase hybridized with FITC-labeled whole chromosome probes to #1, #2, and #4 and counterstained with PI. A reciprocal translocation between one #4 and another chromosome is visible as two bi-color chromosomes labeled “4t”. This false color digital image was produced from a two color digitization of red and green fluorescence.
a b FIG. 2. A 512 x 512 subset of a digitized field. This has been background corrected, and for clarity the
pixel values have been inverted, with the effect of making the fluorescence image look similar to absorption staining. In (a) the borders of the regions of interest (ROIs) are shown in black, and the borders of objects segmented within each ROI in blue. In (b) clusters of objects of suitable size to be metaphase chromosomes have been outlined in red.
FLUORESCENCE METAPHASE FINDER 9
to score in homogeneously stained metaphases (lo), and very time-consuming if scored in G-banded mate- rial (12). However, they are of great interest since the frequency of stable aberrations does not decline signif- icantly over time after the generating exposure event, unlike the unstable dicentrics and acentric fragments, which in peripheral blood have a half-life of -3 years (2). Thus the new test makes possible the scoring, in substantial populations, of chromosome damage result- ing either from a historical event, or from chronic low level exposure. The only problem is that since only a fraction of the genome is painted, only a portion of all the rearrangements will be visible ( la) , and as a con- sequence more cells need to be scored than would oth- erwise be the case.
Previous experience with metaphase finding of ab- sorption stained chromosomes (5, 8, 18) would suggest that a fluorescence metaphase finder would be equally valuable as an aid to visual scoring of painted chromo- some aberrations; perhaps relatively more valuable since the cell population is larger but the analysis time per cell shorter than for the other analyses cited. Al- though metaphase finding in brightfield absorption stained preparations is a solved problem and several commercial machines are available ( l l ) , this is not the case with fluorescently stained metaphase prepara- tions. Fluorescence image analysis presents, if not fun- damental new problems, certainly some engineering challenges. These are mostly caused by three factors. First, the low light level available results in a require- ment for high aperture lenses and cameras that can integrate light for a period from a few tenths of a sec- ond to a few seconds. This limits the scanning speed that can be achieved, and in particular poses a problem for efficient automatic focusing. Second, fluorescence slides frequently display a bright intercellular back- ground. This is particularly noticeable with in situ hy- bridized preparations, where the contrast between bright metaphases and background may be as little as 3:l when excitation specific to 4',6-diamidino-2-phe- nylindole (DAPI) counterstain is employed. The third problem results from the relatively uneven illumina- tion that typically occurs with epi-illumination com- pared with brightfield illumination; combined with the bright background this makes image correction imper- ative.
Almost all commercial metaphase finders, and cer- tainly the faster ones, have been based on special pur- pose hardware (3, 7, 11, 19, 20). With the rapid reduc- tion in the price of increasingly powerful personal computers and scientific workstations, coupled with the absolute limit on scanning speed and data acquisi- tion rate implied by the need to integrate the fluores- cence signal for at least a substantial fraction of a sec- ond, the necessity of using special hardware rather than a general purpose computer is no longer evident. By using a general purpose computer and an operating system common to more than one manufacturer, it is possible immediately to take advantage of the frequent
increases in processing speed or reductions in price of such equipment. Vrolijk (22) has recently reported the development of a fluorescence metaphase finder that uses no special hardware; it was based on a fast model of Apple Macintosh computer. This report describes a fluorescence metaphase finder that was based on a general purpose scientific workstation, programmed entirely in a high level programming language, and running under the Unix operating system. The true and false positive rates of metaphase detection have been measured on a variety of routinely prepared slides, and the value of the finder and consequent rapid retrieval capability has been measured in a scoring experiment that required the visual analysis of many hundreds of metaphases per sample.
MATERIAL AND METHODS Metaphase Preparation, Hybridization,
and Staining Initial studies of the performance of the metaphase
finder were made on a variety of preparations of hu- man peripheral blood lymphocytes, including some pre- pared in our laboratories and some kindly supplied by colleagues at the Lawrence Livermore (CA) National Laboratory. All the slides used in the formal testing reported here had the following properties in common: the slides were hybridized with a t least one whole chro- mosome probe labeled with fluorescein isothiocyanate (FITC) and counterstained with DAPI. Several slides were also counterstained with propidium iodide (PI).
For the visual vs. machine-aided scoring experiment, peripheral blood lymphocytes from hospital X-ray tech- nicians were stimulated with PHA and cultured for 48 h. Colcemid was added for the last 4 h. Cells were treated hypotonically with 75 mM potassium chloride and fixed with 3: l methano1:acetic acid. Slides were prepared with two cell suspension droplets, each typi- cally -15 mm diameter. Slides were hybridized with FITC-labeled whole chromosome probes for chromo- somes #1, #2, and #4 (12). These produced clearly dis- tinguishable labels in normal cells: chromosome #1 was less brightly stained at the centromere heterochro- matic region and at pter than elsewhere, chromosome #2 showed less pronounced but similarly characteristic shading, whereas chromosome #4 was uniformly bright (Fig. 1). These differences in the labeling pat- tern allowed easy verification of each chromosome's presence in normal cells and identification of which chromosomes were involved in rearrangements.
One droplet on each slide (chosen at random) was subsequently used for metaphase finding followed by scoring of machine-found cells; the other was used for purely visual scoring.
Instrumentation Similar multicolor automated fluorescence micro-
scope systems were constructed at MRC Human Genet- ics Unit (HGU) and a t Lawrence Berkeley Laboratory (LBL) from commercially available hardware. These
10 PIPER ET AL.
systems were intended for a variety of uses in addition to metaphase finding; the following equipment was rel- evant to the metaphase finding application:
cient image analysis procedures (14). In order to achieve a reasonable scanning speed, the software was designed to run as multiple parallel processes (16). This permits image processing of previously digitized
An epi-illumination fluorescence microscope (Zeiss [C. Zeiss, Oberkochen] Axioplan at LBL, Leitz [Le- ica, Wetzlar] Ergolux at HGU) was equipped with a x 10, 0.5NA fluorescence lens for scanning (Zeiss Fluar at LBL, Nikon Fluor at HGU) and a variety of higher magnification lenses for review and analysis. A computer-controlled microscope stage was pur- chased either from Ludl Electronic Products (at LBL) or MetaSystems GmbH (at HGU). For automatic scanning, single bandpass epi-illumi- nation filter sets of appropriate wavelengths were provided as follows. At LBL these were, for DAPI, a Zeiss ultraviolet excitation set (H365 + FT395 + LP397), and for PI, a Zeiss green excitation set (H546 + FT580 + LP590). The equivalents at HGU were, for DAPI, a Leitz “A” filter block, and for PI, a Leitz “N2” block. For review and analysis, in each case a filter set specific to FITC fluorescence was available. At HGU, review was assisted by computer controlled changing of the epi-illumination filter sets, whereas at, LBL a multibandpass filter set and separate computer controlled excitation filter wheel was also available for review. A scientific CCD camera, the MicroImager 1400 (Xillix Technologies Corp. Vancouver, B.C.), was used in both cases. This could acquire a field of 1,340 by 1,034 pixels and could integrate light for several seconds without significant dark current. The large field was important for obtaining an acceptable scan- ning speed. With the x 10 objective, the pixel spac- ing at the specimen was -0.68 Fm. For camera interface, we used either an Access Dy- namics DCl/DM11 VSB-bus DMA interface to a framestore that was memory-mapped into the work- station’s memory space via a VME-bus interface (at LBL), or a VME-bus DMA interface directly into the workstation’s main memory (at HGU). A Sun Microsystems (Sunnyvale, Ca) SPARCstation IPX workstation running the Unix operating system and the X Window system was used a t both LBL and HGU. At HGU i t was equipped with a Solflower S-bus to VME-bus converter, whereas a t LBL a Per- formance Technologies SBS915 S-bus to VME-bus converter was used.
Recently, a system essentially similar to the one at LBL but based on a Nikon Microphot SA fluorescence microscope has been constructed and brought into use at the Division of Molecular Cytometry, UCSF. The results reported here were all obtained using the LBL machine.
Metaphase Finding Software Software was written in the C programming lan-
guage and was based on an established library of effi-
frames to overlap with stage movement and image ac- quisition of the next frame. It also will port easily to multiprocessor machines for increased performance.
Auto focus. Focus was maintained by precomput- ing a focus map from which the focus position of the current stage position was found by interpolation. At preset regular iytervals during scanning, the predicted focus level was compared with a newly measured focus level. Any difference was assumed to be due to drift in the Z axis and resulted in an appropriate offset being added to the focus map.
In order to increase focusing speed, the camera was operated in “binning” mode, in which each digitized pixel was derived from a 2 x 2 patch of pixels on the camera CCD. As a result, the number of image pixels was reduced by a factor of four, the light sensitivity increased by the same factor, and the integration time during focusing was therefore reduced by a factor of four.
The contrast function used was Vollath’s F, (21).
the digitized frame, where g,,yl is the pixel intensity a t position (x,y) and s is a scaling parameter. The function thus computes the total signed gradient of the image in the horizontal direction, weighted by local intensities. Vollath (21) showed that this particular focus function has good immunity to noise in the image. In this ap- plication, s was chosen to be 3 (binned) pixels. With a x 10 objective lens, this function proved to give good focus determination even from fields with little and/or faint foreground material.
Region of interest determination. Images cap- tured during metaphase finding typically consisted al- most entirely of intercellular background. Even on slides with a high cell density, it was usually the case that < 10% of the pixels belonged to non-background objects. To increase the speed of image analysis, a “re- gion of interest” (ROI) approach was adopted (7,22), in which a preliminary analysis in a subsampled version of the image was used to detect above-background re- gions for more detailed analysis a t full resolution. The subsampled image consisted of every fourth pixel in every fourth row of the original image, i.e., the number of pixels was 1116th that of the full resolution image.
Background correction. Whereas the intercellu- lar backgrounds of both PI counterstained and unhy- bridized DAPI stained slides were usually dark, it was found that the background of hybridized slides was typ- ically very bright when visualized with DAPI counter- stain, so that the relative contrast between metaphase and background pixels was usually < 3:l. This back- ground arose from several sources. The camera dark current, the analog to digital convertor DC offset, and the autofluorescence of the objective (which was found to be substantial with ultraviolet excitation), taken to-
This sums the expression g,,,yl. @(x+s,yl - gjx+2sJ over
FLUORESCENCE METAPHASE FINDER 11
gether contributed a signal level which varied across the image but was effectively constant from one image to the next. However, the major contribution to the background signal was nonspecific fluorescence of the intercellular background.
Unlike previous experience with brightfield absorp- tion stained metaphase preparations, it was found that the average brightness both of the cellular components and of the intercellular background varied from field to field. A number of factors may contribute to this, e.g., uneven staining, fading of fields that have been exam- ined previously, and an increase in glare near the edges of the cover slip. We note that none of these nec- essarily make a slide less suitable for visual analysis. Whereas it might be possible in some circumstances to reduce the variation by more careful attention to slide preparation and storage, it is important that a meta- phase finder be capable of working with the quality of material in general day-to-day use in a laboratory; a metaphase finder that requires special attention to preparation will not win wide acceptance.
Because the background was not constant across an individual field, background correction was required before a global threshold could be used for segmenta- tion. During a preliminary calibration phase, a back- ground image was obtained by scanning a few (typi- cally five) adjacent fields. The value of each pixel of the background correction image was defined to be the minimum value of the corresponding pixels in the set of adjacent images. The mean value of the background correction image was also computed.
Because the background varied between fields, use of a single background compensation image did not pro- vide a complete solution. During scanning, each pixel of an image was corrected, first by subtracting the cor- responding pixel of the background image, and second, the difference between the mean of the background im- age and the mean of the current image, plus a small positive constant, was added to each pixel. The second stage provided approximate “background following” to take account of background brightness that varied across the slide and prevented images with fainter background than the background correction image from becoming negative after correction.
Both the subsampled image of the full field and the full-resolution ROIs were corrected in this fashion be- fore further analysis.
Thresholding and segmentation. Because of the observed variability in the fluorescence brightness of the cellular components mentioned previously, the use of a constant threshold for the entire slide was not ex- pected to be adequate, even after background correc- tion. In this respect, fluorescence metaphase finding differs substantially from absorption stained meta- phase finding, where a single threshold sufficed for a whole slide (20).
It has been our common experience with microscope slide preparations that an appropriate threshold T for segmentation may be expressed as T = B, + 0, where
B , is the upper limit of the distribution of background values, and 0 is a positive offset. B , is simple to mea- sure from the pixel intensity histogram and, except in extreme circumstances, is independent of the amount of foreground (cellular) material. Finding an appropri- ate value for 0 is not so simple; it appears to depend in particular on the range of brightness levels. In the case of fluorescence metaphase finding, this range may vary from field to field. A compounding problem is that several popular techniques for choosing the threshold from the intensity histogram will produce different threshold values if the image has relatively more or less foreground material (examples are methods based on histogram entropy measures, e.g. (9), and the iso- data method (17)); this is clearly unacceptable in our application.
In order to avoid these problems, a threshold for a particular image was computed by a two-stage process, based on the subsampled version of the image. First, an initial threshold was chosen by finding the clearly rec- ognizable background peak in the histogram of pixel intensities, and its upper limit B,. Retaining only the pixels brighter than B , typically resulted in a subset of the image that contained the interphase nuclei and metaphases surrounded by some of the neighboring background. Thus this first segmentation stage en- sured that the subsequent stage was not dominated by intercellular background, but instead processed image regions composed of a representative mix of both fore- ground and adjacent background material.
In the second stage, a higher second threshold was chosen from a two-dimensional histogram of pixel gray value vs. gradient at the pixel, measured over the re- gion defined by the first threshold. The second thresh- old was chosen to be the mean gray level of those pixels whose gradient exceeded the mean gradient level, i.e., there was a bias in favor of a value at which pixels had a relatively high gradient, thus ensuring that the cho- sen threshold did indeed correspond to the edges of ob- jects in the image.
Since abrupt variations in fluorescence brightness across the slide do not occur, it was expected that ad- jacent fields should have similar threshold values. Thus a running average threshold value was calcu- lated and used in preference to the proposed threshold computed on the current field. The running average threshold was weighted by the number of pixels above B , in each successive image, but with relatively higher weight given to more recent images so that it remained sensitive to local differences in threshold level. This had the desirable effects of preventing wild fluctuation in threshold levels on sparse images, and maintaining the threshold value across empty regions of the slide.
Finally, the ROIs were obtained by dilating the above threshold part of the subsampled image, finding the connected components, and magnifying them to correspond to the full resolution image (Fig. 2a). Each ROI then defined a connected region of the full resolu- tion image. In turn, each such region was background-
12 PIPER ET AL.
corrected and segmented using the same threshold as the subsampled image. The result in each ROI was a set of connected components of the image (Fig. 2a); these formed the basis of the search for metaphases.
Detection of metaphases. Each set of connected components was divided into “limbs” (19), i.e., non- branching subsets. Limbs that were either smaller or larger than preset limits were discarded. Remaining (chromosome-sized) limbs were grouped into clusters by single-link clustering based on Euclidean distance between limbs (Fig. 2b). In order to join a cluster, a limb had to be closer than a distance threshold to a t least one existing member of the cluster.
The limb size limits and the cluster distance thresh- old, along with other parameters of the system, can be adjusted by the user. However, after an initial period of experimentation, the majority of such parameters have acquired values that appear to be generally applicable, and they are now rarely, if ever, varied.
Features were measured for each cluster: in partic- ular, the number of limbs, their total area, the area and aspect ratio of the convex hull enclosing the cluster, and the total intensity of the set of limbs. A simple classifier compared each feature value against a set of limits determined interactively in a training experi- ment. If every feature value was within the specified limits then the cluster was classified as a metaphase and its location recorded. Feature limits and other pa- rameters could be set by the user. However, in the ex- periments reported here, only camera integration time and the scan area were adjusted.
Following completion of the scan, the list of meta- phase coordinates and feature measurements was stored to disc. This permitted the review or analysis phases to be delayed to another time, if desired.
Metaphase Finder Performance Measurements False positive and false negative rates were deter-
mined by scanning -0.6 cm2 on each of nine hybridized slides counterstained with DAPI. The program paused after acquiring and displaying each field scanned, so that the operator could determine the true number of metaphases in the scanned field (viewing i t through the microscope eyepieces in cases of doubt), and relate these to the program’s output. Metaphases judged un- suitable for subsequent scoring by reason of being in- complete, double, or poorly spread, were recorded as a separate category from scorable metaphases. False pos- itives were also recorded. Three of the slides had also been counterstained with PI; these were scanned over the same area as in DAPI in order to compare the re- sults from scanning in the two different counterstains.
Review of Located Metaphases Following metaphase finding, the cytogeneticist
could review the metaphases found and a variety of options were available to assist in this phase. Each metaphase found by the program was relocated, cen- tered under the objective. The initial relocation posi-
tion was not always central in the field of view, partic- ularly if the slide had been removed from the stage in the interim, or if a different power objective was se- lected. Once the first metaphase had been relocated and its centering adjusted, an offset was automatically computed so that subsequent metaphases were cor- rectly centered.
An advantage of using a fluorescence microscope was that metaphases could be reviewed with a different excitation wavelength to that used for initial scanning, e.g., metaphases could be reviewed while looking di- rectly a t an FITC labeled hybridization. The reviewing options included image grabbing and screen display of relocated metaphases, in which case the digital image contrast was adjusted to fill the dynamic range of the display, and autofocus (necessary in particular if re- viewing at high magnification). The digital image fre- quently proved valuable in interpreting the configura- tion of particular metaphases.
Metaphases could be presented in the order in which they were found, or alternatively in order of a quality measure computed from features of each metaphase (20). The quality computation could be trained in a preliminary experiment to suit each cytogeneticist’s preferences. A visually determined classification could be attached to each metaphase; this was used, for ex- ample, in the translocation scoring experiment de- scribed below.
Scoring Translocations in Painted Chromosomes in Machine Located Metaphases
Slides were scanned for metaphases over a rectangu- lar region that covered one of the two-cell suspension droplets. The metaphases found were reviewed at higher power ( x 63 oil). The epi-illumination filter block was changed from the single-bandpass block spe- cific to the counterstain used for metaphase search, to a multibandpass block (“Pinkel-1”, Chroma Technol- ogy). Computer-controlled changing of the excitation filter then allowed easy and rapid switching between hybridization and counterstain labels, assisting in ac- curate visual analysis. Note that the criterion for a metaphase to be judged “scorable” in this experiment was different than in the performance measurement experiment described above, since here it included evaluation of the quality of the in situ hybridization signal. Furthermore, since only three chromosomes were labeled for translocation scoring, the underlying frequency of cells with translocations should be calcu- lated as described previously (12).
A numerical keypad was coded to provide the oper- ator with the following facilities on single keystrokes:
Classify as a scorable, normal metaphase; move to
0 Classify as an unscorable metaphase; move to the
Classify as a nonmetaphase (false positive); move to
the next metaphase.
next metaphase.
the next metaphase.
FLUORESCENCE METAPHASE FINDER 13
Move to the previous metaphase (in order to recon- sider decision taken). Make one of a set of precoded classifications of ab- normality (translocation, dicentric, etc); move to the next metaphase. Change excitation filter Reset focus offset
The totals of metaphases in each category were accu- mulated and the time for review measured. The meta- phase list coordinate was annotated with the classifi- cation and stored on disk, so that classifications could be reviewed at a later date.
Thirteen slides prepared from five cases were scored in the fashion described, in order to determine to what extent a metaphase finder would assist in a larger study.
Unaided Visual Scoring of Translocations The other cell droplet on each of the slides scored in
the manner described above was used for a purely vi- sual control scoring. Metaphases were both found and then analysed a t x 63 (oil) using the same criteria as described above, on a microscope with no automation components. Given that peripheral blood lymphocyte preparations are usually rich in metaphases and that using x 63 for finding followed immediately by analy- sis of each suitable metaphase was the standard proce- dure in the laboratory prior to the current experiment, this method was expected to be faster than using x 10 for finding followed by x 63 for analysis.
RESULTS Metaphase Finder Performance Measurements Typically it required -5 min to load the slide, choose
the appropriate integration time, specify the scan area, and make any other desired adjustments to operating parameters, and another 45 s to compute the focus map and the background correction image. Following this initial setup time, the scanning speed achieved with the material used here was 0.16 mm2/s. The time taken for review and scoring was variable depending on the nature of the analysis, but the time to relocate and refocus each metaphase and then proceed to the next as fast as possible averaged 2.1s per metaphase.
Table 1 shows the true positive, false negative, and false positive frequencies measured on nine DAPI- counterstained slides, together with comparable fre- quencies for the same areas of the three slides in the set that also had PI-counterstaining. The ratio of false pos- itives to all positives (including both DAPI and PI scores for the slides scanned in both) was 9.3%. This figure has proved to be dependent on the number of metaphases and the quantity of debris on the slide. Thus the range of the false positive rate varied consid- erably among the preparations in Table 1 (0-25%), with scanning in PI proving to be consistently better.
The false negative frequency (missed metaphases/to- tal metaphases) was 20% for the entire data set in Ta-
ble 1, with a range among the slides of 7.5-53% false negatives. However, if the false negatives are more stringently defined (as missed scorable metaphases/to- tal scorable metaphases) then the frequency was 8.7% (range 1.7%-25%). As with the false positive rate, PI staining yielded lower false negative rates in general. In summary, 80% of all metaphases and 91% of those metaphases judged to be scorable were found by the machine.
Scoring Translocations in Painted Chromosomes in Machine Located Metaphases and Unaided
Visual Scoring of Translocations Table 2 shows that the speed at which metaphases
presented by the machine could be scored averaged more than three times faster than purely visual scor- ing. If -5 min of operator time, needed to set up the metaphase search on each slide, is also taken into ac- count, the relative speed increase was still a factor of more than two.
As was observed above for the accuracy of metaphase finding, the time per metaphase for review, whether purely visual or machine assisted, was quite variable depending upon the sample preparation. For example, in sample X-2 there was a high concentration of closely positioned scorable metaphases, resulting in particu- larly efficient visual scoring. In addition, the frequency of nonmetaphases and unscorable metaphases identi- fied by the machine had a significant influence on the speed of machine assisted analysis, so that sample X-15, which had an unusually high frequency of un- scorable events, was analyzed more slowly than the other four samples.
Table 2 also shows a tendency for more scorable metaphases to be located by machine than were de- tected by the visual scan (at x63) of the alternative droplet, even though the areas scanned by each tech- nique were similar in size; on average, the machine identified almost 50% more scorable metaphases.
Within the limitation of the small sample of slides reported here, there was no apparent difference in the frequency of translocations found by the two tech- niques.
DISCUSSION Fluorescence metaphase finding is one application of
the general purpose automated fluorescence micro- scopes that we have constructed. The metaphase finder clearly works well, if rather slowly by comparison with some specialized absorption metaphase finders.
Korthof and Carothers (11) reported extensive inde- pendent tests of four commercial absorption metaphase finders and found a mean false positive rate of 16% averaged over all four machines and eight slides that were relatively “rich” in metaphases (two each of lym- phocyte, fibroblast, amnion, and long-term chorionic villus cultures). Unfortunately, they did not measure the false negative rate on these slides. On four “sparse” slides (two each of bone marrow and short-term chori-
“Nine DAPI stained slides, together with the performance on the same area scanned in PI counterstain for three of the slides. The slight discrepancies between cell totals in PI and DAPI in the three slides scanned in both counterstains may be explained by (1) the machine not scanning precisely identically positioned fields, and (2) the expected absence of perfect repeat- ability of the cytogeneticist’s visual scoring of the metaphases.
bMetaphases that were correctly identified by the machine, with the ratings (Scorable and Unscorable) defined by the cytogeneticist.
“Metaphases detected by the cytogeneticist, but not identified by the machine, with the ratings (Scorable and Unscorable) defined by the cytogeneticist.
dFalse positives are items identified as metaphase cells by the machine that were determined by the cytogeneticist not to be metaphase cells.
Table 2 Comparison Between Visual and Machine-Assisted Scoring
“Number of translocations detected. bAll times in this table are in minutes. “Speed is reported in metaphases scored per min. dObjects that were identified as metaphases by the machine, but were determined by the cytogeneticist not to be metaphases
“Ratio of machine assisted speed to visual scoring speed. ‘One cell was found that contained a dicentric chromosome. “Speed reported is the average speed (total metaphases scoredltotal review time). h T ~ o averages of the relative speed are reported. The first figure (3.2) is the average relative speed determined as (average
machine-assisted speeaaverage visual speed). The second figure (4.1) is the unweighted average relative speed, defined as the mean of the relative speeds of all samples.
(false positives).
onic villus samples), the average false positive rate was 65%, whereas the false negative rate over the four ma- chines was 41%. Vrolijk (22) recently reported a false positive rate of 7.4% and a false negative rate of 12.7% on six “rich” DAPI counterstained fluorescence slides. The nine slides used in our tests were all “rich”; the
false positive and false negative rates that we obtained (Table 1: overall, 9.3% and 20.0%, respectively) indi- cate that our system’s performance is relatively good, particularly since the false negative rate of “scorable” metaphases was only 8.7%. The mean false positive rate of 7.7% obtained on a different set of 13 slides used
FLUORESCENCE METAPHASE FINDER 15
in the aberration scoring experiment (Table 2) is sim- ilarly acceptable.
Korthof and Carothers (11) also measured both the scanning speed and review times of the four systems tested. The three systems that were based, like ours, on frame cameras and stop-start scanning had speeds in the range 0.21 mm2/s0.37 mm2/s, averaged over 12 test slides. Vrolijk’s fluorescence metaphase finder had a speed of 0.22 mm2/s (22). Our reported speed of 0.16 mm2/s is thus relatively slow. However, making use of the “binning” capability of the CCD camera (thus ef- fectively scanning at 1.35 pm pixel spacing) increased the scanning speed to 0.31 mm2/s, a t the expense of some loss of accuracy (data not presented). The speeds reported here refer to the LBL machine; our experience has been that (largely on account of its different cam- era interface) the HGU machine is somewhat faster (data not presented).
Currently, the scanning speed is processor-limited on the SPARCstation IPX. Most of the computational load comes from the low-level processing concerned with background correction, background following, and ini- tial thresholding. These are simple operations that could be carried out by hardware, although this would be both more expensive and less flexible. The ultimate constraint on speed is the time taken to integrate suf- ficient light while the stage is stationary, added to the time for the stage to move from one frame to the next and then settle. With our existing stage and camera, together these imply a limit of (at best) one frame per s (or -0.6 mm2/s). If a four times more powerful pro- cessor were to be employed (or a multiprocessor, which the multiprocess structure of the system can exploit without modification), then this limiting speed would be achieved. Since the experiments reported here were performed, such machines have become generally available a t a reasonable price (e.g., in the Sun Mi- crosystems SPARCstation 10 range). We therefore see no benefit in designing (or buying) specialized image processing hardware.
Review time per metaphase depends on the nature of the analysis. We measured the minimum time taken to relocate at high power and refocus (if necessary) a set of metaphases, if no other evaluation was performed; this averaged 2.1s per metaphase, very similar to the over- all average of 2.0s per metaphase for the four meta- phase finders tested by Korthof and Carothers (1 l).
Although the finder works well with either PI or DAPI counterstains, we have found that the typically bright DAPI background of in situ hybridized slides makes the setting of parameters such as camera inte- gration time much more critical (since the experiments reported here were performed, automatic setting of the integration time has been added to the system, with the additional benefit of a reduction in the setup time per slide). However, the DAPI banding was distinct enough to confirm the identify of each of the three painted chromosomes in a metaphase and was there- fore preferred to PI as a counterstain for aberration
scoring. Reduction of the intercellular background is recommended as a good target for those who make metaphase preparations.
The background correction and following method that we have used is clearly only an approximation to the full correction that might be employed, e.g., by compensating for illumination shading by a map ob- tained using a uniformly fluorescent standard slide, such as described by Galbraith et al. (6). However, we have found that illumination shading is less of a prob- lem than the variation from field to field caused by nonuniform fluorescence intensity across the slide, and in these circumstances even full correction would still not remove the desirability of a different threshold for each field. It would however require a further calibra- tion step involving a separate slide and cost additional processing time.
The autofocus system usually worked well so long as there was adequate material in the field of view on which to focus. However, brightly fluorescent dust on top of the coverslip was able to confuse the focusing program if the metaphase preparation was itself any- thing less than brightly stained. It proved important always to clean the coverslip carefully before scanning.
Turning to the aberration scoring experiment, whereas the visual scoring speed varied between 2 and 9 scorable metaphases per min, the machine-assisted scoring speeds lay between 8 and 16 scorable meta- phases per min. The average speed increase of more than three is in close agreement with that previously published for absorption stained chromosome aberra- tion scoring ( 5 ) . If we assume an average scan setup time of 5 min per slide, then machine scoring was still overall more than twice as fast as purely visual scor- ing. Since search setup time is independent of the area scanned or the density of material on the slide, the optimum way to exploit machine scanning for aberra- tion scoring would be to adjust sample preparation to give a high density of metaphases on the slide, spread the cell suspension over as large an area of the slide as possible, and scan it all at once, all in order to minimize the relative cost associated with scan setup. We also expect that the speed of fluorescence metaphase scor- ing could be maintained at higher rates (10-20 meta- phases/min) by use of an effective metaphase quality ranking algorithm, so that the cytogeneticist would be- gin by scoring the “best” metaphases and discontinue when unscorable or nonmetaphases became predomi- nant. A ranking algorithm has been developed for the system described here, but its benefit in aberration scoring has not yet been measured; in the present ex- periments all objects found by the machine were either scored or rejected.
An unexpected result visible in Table 2 was that on average the machine found -50% more scorable meta- phases than the cytogeneticist did on the alternate cell droplet. The first three slides were therefore also scanned by machine on the same droplet as the cytoge- neticist had scored, and this confirmed that the ma-
16 PIPER ET AL.
chine did indeed find many more scorable metaphases (data not presented). The explanation is probably that, following previous practice in our laboratory, visual search for metaphases used the same x 63 objective as was used for the analysis. Maintaining a uniform vi- sual search pattern with a high power objective is clearly not easy, and some scorable metaphases are likely to be missed. However, using a separate ~ 1 0 search phase would probably find more metaphases overall but would certainly increase dramatically the time per metaphase analyzed on account of the time needed to record (manually) the metaphase locations and subsequently relocate them at ~ 6 3 . It is also worth pointing out that the ~ 1 0 visual image in a fluorescence microscope is many times dimmer than that seen at x 63, which is the reverse of the experience with transmission microscopy. Shafer (18) reported that automatic metaphase finding reduced interopera- tor variability in SCE scoring, most probably because each cytogeneticist was constrained to analyse the same metaphases. The comparison of the number of metaphases found by visual search and by machine scanning shows that a similar gain in reproducibility is occurring with this system. Furthermore, the increased yield of scorable metaphases would give significant benefit when performing aberration scoring on small samples.
Whereas the metaphase finder makes the scoring of rare cytogenetic aberrations substantially faster, it does not completely remove the tedium of the task, and analysing a large population of individuals still re- quires a huge amount of a skilled cytogeneticist’s time. At the rate of 10 metaphases per min and given the relative sensitivity obtained by painting just three chromosomes (-30%) (12), it would require -1 h to perform a translocation frequency enumeration of the equivalent of 200 cells. This is much faster than was previously possible, but it amounts to only five whole cases per cytogeneticist per day, still an unacceptably slow analysis rate for large population studies. We have previously constructed a fully automatic pre- screening system for absorption stained dicentric chro- mosomes (1) and are currently developing a prescreen- ing system for detecting rearrangements of painted chromosomes of the sort scored in the experiment re- ported here (15), in the expectation that this work will increase the analysis speed and reduce its cost still further.
ACKNOWLEDGMENTS We are grateful to our colleagues J. Lucas, J. Tucker,
D. Lee, and M. Ramsay at Lawrence Livermore Na- tional Laboratory for providing metaphase prepara- tions that extended the range of material available for testing the metaphase finder performance.
LITERATURE CITED 1. Bayley R, Carothers A, Chen X, Farrow S, Gordon J, J i L, Piper
J, Rutovitz D, Stark M, Wald N: Radiation dosimetry by auto-
2.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
matic image analysis of dicentric chromosomes. Mutat Res 253: 223-235, 1991. Bender MA, Awa AA, Brooks AL, Evans HJ, Groer PJ, Littlefield LG, Pereira C, Preston RJ, Wachholz BW: Current status of cy- togenetic procedures to detect and quantify exposures to radia- tion. Mutat Res 196:103-159, 1988. Castleman K R The PSI automatic metaphase finder. J Radiat Res 33(suppl):124-128, 1992. Cremer T, Lichter P, Borden J , Ward DC, Manuelidis L: Detection of chromosome aberrations in metaphase and interphase tumor cells by in situ hybridization using chromosome-specific library probes. Hum Genet 80:235-246, 1988. Finnon P, Lloyd DC, Edwards AA: An assessment of the meta- phase finding capability of the Cytoscan 110. Mutat Res 164:lOl- 108,1986. Galbraith W, Wagner MCE, Chao J , Abaza M, Ernst LA, Nederlof MA, Hartsock RJ, Taylor DL, Waggoner A: Imaging cytometry by multiparameter fluorescence. Cytometry 12:579-596, 1991. Graham J, Pycock D: Automation of routine clinical chromosome analysis. 11: Metaphase finding. Anal Quant Cytol Histol 9:391- 397, 1987. Green DK, Bayley R, Rutovitz D: A cytogeneticist’s microscope. Microscopia Acta 79:237-245, 1977. Kapur JN, Sahoo PK, Wong AKC: A new method for gray-level picture thresholding using the entropy of the histogram. Comp Vision Graph Image Proc 29:273-285, 1985. Kleinerman RA, Littlefield LG, Tarone RE, Machado SG, Blett- ner M, Peters LJ, Boice JD: Chromosome aberrations in periph- eral lymphocytes and radiation dose to active bone marrow in patients treated for cancer of the cervix. Radiation Res 119:176- 190, 1989. Korthof G, Carothers AD: Tests of performance of four semi-au- tomatic metaphase-finding and karyotyping systems. Clin Genet 40:441-451, 1991. Lucas JN, Awa A, Straume T, Poggensee M, Kodama Y, Nakano M, Ohtaki K, Weier H-U, Pinkel D, Gray J, Littlefield G: Rapid translocation frequency analysis in humans decades after expo- sure to ionizing radiation. Int J Radiat Biol 6253-63, 1992. Pinkel D, Landegent J , Collins C, Fuscoe J , Segraves R, Lucas J, Gray J : Fluorescence in situ hybridization with human chromo- some-specific libraries: Detection of trisomy 21 and translocations of chromosome 4. Proc Natl Acad Sci USA 859138-9142, 1988. Piper J , Rutovitz D: Data structures for image processing in a C language and Unix environment. Patt Recogn Letts 3:119-129, 1985. Piper J, Bayley R, Boyle S, Fantes JA, Green DK, Gordon J, Hill W, J i L, Malloy P, Perry P, Rutovitz D, Stark M, Whale D: Au- tomatic aberration scoring using whole chromosome FISH. Cy- tometry supp1.6:56, 1993. Poole I, Piper J , Rutovitz D, Stark M: Cytoline-software design. Internal Report IN94-002, MRC Human Genetics Unit, Edin- burgh, UK, 1994. Ridler TW, Calvard S: Picture thresholding using an iterative selection method. IEEE Trans Syst Man Cybern SMC-8:630-632, 1978. Shafer DA, Mandelberg KI, Falek A: Computer automation of metaphase finding, sister chromatid exchange, and chromosome damage analysis. In: Chemical Mutagens: Principles and Meth- ods for Their Detection, Vol. 10, de Serres FJ (ed). Plenum Press, New York, 1986, pp. 357-380. Shippey G, Bayley RJH, Farrow A S , Rutovitz D, Tucker JH: A fast interval processor. Patt Recogn 14:345-356, 1981. Shippey G, Carothers AD, Gordon J : Operation and performance of an automatic metaphase finder based on the MRC Fast Inter- val Processor. J Histochem Cytochem 34:1245-1252, 1986. Vollath D: The influence of the scene parameters and of noise on the behaviour of automatic focusing algorithms. J Microscopy 151:133-146, 1988. Vrolijk H: Automation of chromosome analysis, PhD thesis, Uni- versity of Leiden, The Netherlands, 1993.
