pare effects of preprocessing algorithms used for auto- ... using a semi-automated analysis program (17). ... trary highpass filter and also a more involved tech-.
0 1990 Wiley-Liss, Inc.
Cytometry 11:40-50 (1990)
Stylized Chromosome Images’ Alice 0. Martin,2Barry S. Isenstein, Cecil W. Thomas, Mark S. Rzeszotarski, and Walter E. Johnson Department of Obstetrics and Gynecology, Northwestern University Medical School, Chicago, Illinois 60611 (A.O.M.);Mercury Computer Systems, Inc., Lowell, Massachusetts 01854-3608 (B.S.I.); Department of Biomedical Engineering (C.W.T., M.S.R.) and Genetics Center (W.E.J.),Case Western Reserve University, Cleveland, Ohio 44106 Received for publication April 6, 1989; accepted July 26, 1989
Stylized chromosome images 1) serve as a format to test effects of preprocessing algorithms used in automated karyotyping; 2) enhance the ability of humans to perform quantitative analysis of chromosomal aberrations; 3) provide an alternative format for karyotype hard copies produced by automated systems. Stylized chromosomes are two-dimensional computer-generated images based on information extracted from one-dimensional width and density profiles. These profiles correspond to what cytogeneticists observe through the microscope as the shape and banding patterns of stained chromosomes. Stylized presentation sharpens chromosome band boundaries and perimeters, reduces “noise,” and enhances gray level variations, which are difficult to distinguish by humans on photographic or computer generated karyotypes. Karyotyping accuracy using stylized
Images of chromosome banding patterns and chromosome perimeters are not distinct because there is no sharp dcmarcation of chromosomal from extrachromosoma1 material in the living or stained nucleus. There are also no sharp distinctions between the horizontal “bands,” because these are visual manifestations both of underlying heterogeneity in chromosome structure and variation in chromosomal preparation and staining techniques. Despite these biological and laboratory realities, cytogeneticists prefer distinct band appearances and attempt‘ to sharpen band boundaries by perfecting laboratory preparation and staining methods, and photographic techniques. Manufacturers of automated systems also strive to produce sharper images by increasing resolution and by applying algorithms to enhance contrast of digital images. (Information from a real chromosome is converted to a set of numbers proportional to gray levels.) If karyotypes are to be used
images was used to detect difficult areas for automated chromosome identification. Landmark bands sufficient to classify chromosomes were identified; shapes of chromosomes reflected in width profiles were said to aid classification. A two-step automated karyotyping strategy proposed is: 1) classify chromosomes by landmarks, minimum information needed for identification; 2) subsequently employ the full banding pattern with maximum resolution to detect aberrations. Stylized images of abnormal chromosomes have potential for testing hypothesis regarding breakpoints and quantitative analysis, but improvements are needed in homologue normalization and definition of termini of chromosomes. Key terms: Chromosomes, automated karyotyping, chromosomal aberrations, integrated density profiles
for diagnostic purposes, care must be taken not to sacrifice accuracy for beauty. However, if the purpose of karyotype preparation is not for analysis but to send records to patients andlor physicians, aesthetic considerations may predominate over diagnostic ones. Ideally both should be achieved. We produced stylized chromosome images with distinct band boundaries and perimeters, capable of exhibiting band resolution varying from the minimum landmarks required for classification to the maximum
’This work was supported in part by NIH-GM-26499, and the Whitaker Foundation.
NIH-HD-8-2855,
‘Alice 0. Martin, Ph.D., is now a t Arnold, White and Durkee, 800 Quaker Towers, 321 N. Clark, Chicago, IL 60610. Address reprint requests there.
41
STYLIZEI)CHROMOSOME IMAC-ES
distinguishable for aberration detection. A stylized chromosome is a two-dimensional computer-generated image based on information extracted from the onedimensional width and density profiles constructed from digital chromosome images. These profiles correspond to what cytogeneticists observe through the microscope as the shape and banding pattern of the stained chromosomes. Presentation in a two-dimensional stylized manner sharpens boundaries and perimeters, reduces “noise,” and enhances gray level variations difficult to distinguish by the human on photographic or hard copy karyotypes. Stylized chromosomes were initially designed to compare effects of preprocessing algorithms used for automated karyotyping. It is generally assumed that all the information necessary for automated karyotyping is contained in the density profiles (4,5).The stylized images represent the same information used by the computer for karyotyping, in a format more similar to that the cytogeneticist has been trained to recognize ( 3 , l O ) . Individual features of density profiles, combinations of features, and manipulation of profiles by various preprocessing algorithms may be selected to generate stylized images. Comparisons of the accuracy of human karyotyping using these images produced in different ways can assess the importance of those features for chromosome classification. Because of the similar appearance of stylized to real chromosomes, we would expect evaluation to be easier than methods comparing density profile changes to see if information is lost (46 3 ) .The additional information required for aberration detection after classification can be determined. Stylized images thus provide a “common ground,” a visual bridge between the efforts of engineers and the desires of cytogeneticists. We compared karyotyping accuracy using stylized versus digital chromosome images to determine whether significant information is retained after profile preprocessing, to determine the gray level resolution, and sampling density required for classification, to explore “landmark band classifiers” (4,141, and to analyze chromosome aberrations.
MEAN -DENSITY PROFILE
INTEGRATED DENSITY PROFILE
9 WIDTH PROFILE
BOUNDARY ANDCENTERLINE
CHROMOSOME NO. I FIG.1. Intermediate computer output from a digital image for a typical chromosome 1. The fitted centerline and centromere are indicated. The tick marks seen on the X and Y axes are equal to 50 samples of 0.05 pm each. The mean density profiles and width profiles are used to generate the stylized chromosomes.
3. Mean density: integrated density profile divided by width profile.
Peaks (valleys) are locations on the density profile whose density is greater (less) than the density values of points a t a distance of two to four samples away from that location in both directions. Peaks represent chromosome bands which are G positive; valleys represent G negative bands. The centromere location is the valley in the inner MATERIALS AND METHODS 80% of the width profile; if there is no minimum, the Production of Stylized Chromosome Images The centerlines of chromosomes were determined by centromere is assigned to the 10% point on the end a cubic-spline curve fitting algorithm (1).Small chro- which has the lowest mean width summed over the mosomes (e.g.,the G-groups) were fitted with a straight- 10% extreme. Four-by-five photographic negatives of human cells line centerline. Chromosomes were excluded if the cenin metaphase were digitized a t a n equivalent resoluterline could not be fit, or gave a n obviously erroneous tion of 22.5 samples per km and 256 gray levels, well fit. In practice, this meant that severely bent chromosomes were excluded. A sample corresponds to 0.05 Fm. above what has been determined as sufficient for auFeature extraction used perpendiculars to the center- tomated karyotyping (3,lS). Profiles were generated line for the compilation of the following profiles (Fig. 1): using a semi-automated analysis program (17). The width profile was used to construct perimeters 1. Width: defined by lengths of the perpendiculars; for the stylized chromosomes. The variation along the a t a single centerline point this is one-half the distance length of the chromosome was determined by the mean density profile. Bands were constructed so there was no between boundaries. 2. Integrated density: sum of the density values variation across their width. Presentation of informaalong a given perpendicular to the centerline. tion extracted from the one-dimensional width and
42
MARTIN ET AL.
density profiles, in a two-dimensional format, takes advantage of human visual processing to sharpen band boundaries.
Preprocessing Density Profiles Three general techniques of preprocessing density profiles to generate stylized chromosomes were employed: grayscale remapping, a n heuristic algorithm, and linear filtering. Grayscale remapping is a process by which the grayscale is literally redefined such that separation between adjacent gray-levels can be “stretched” at the expense of “compressing” another region of the grayscale (2).The grayscale range used was 0 (black) to 255 (white). The remapping transformation table is given by Isenstein ( 3 ) . The parameters can be changed, producing a n almost unlimited number of remapped images. Either global or local remapping is possible. Grayscale remapping has been implemented in similar interactive modes for other medical imaging modalities, e.g., CAT scan and ultrasound (19). A n heuristic algorithm is designed to maximize band contrast. Briefly, it involves finding the first dark band of each end of the chromosome, rescaling the interior portion of the chromosome to maximize contrast by assigning minimum and maximum densities 0 and 255 gray levels, respectively, and then adjusting the end bands for consistent display. The heuristic procedure takes full advantage of the available dynamic range of the display device for the information carrying portion of the density wave form. Linear filtering has as its objective a particular operation, e.g., to enhance all bands. We used a n arbitrary highpass filter and also a more involved technique called inverse filtering to fit incident profile information to a reference model, e.g., that of “landmarks” used by cytogeneticists to classify chromosomes. In experiments using inverse filters, we defined the model to be a minimal karyotypable metaphase, i.e., a metaphase in which chromosomes contain all the necessary “landmark” bands and few extraneous bands. Landmarks are determined by reducing information extracted from the density profiles to generate stylized chromosomes until cytogeneticists could not longer identify the chromosome. Linear filtering caused a poor quality metaphase to have its bands enhanced, while in a n excellent quality metaphase small bands were de-emphasized. The rationale behind this operation was to aid automated karyotyping (chromosome classification) by de-emphasizing variable bands not necessary for karyotyping. We have simplified the computer’s job by removing bands that might cause confusion (i.e., reduced metaphase to metaphase banding variation). However, as detailed band analysis is required for aberration detection, all possible banding information present in incident chromosomes is required. Consequently, after classifying chromosomes in step 1 using a landmark band classifier, small bands would be displayed for subsequent analysis.
Evaluation of Karyotyping Using Stylized Images Our overall system is illustrated in the block diagram of Figure 2. Evaluation of results was based on discussion with cytogeneticists regarding ease of recognition of the chromosomes a t successive steps of data reduction for determination of landmarks, and on trials of karyotyping accuracy using digital versus stylized images prepared from the same metaphases. The chromosomes for this study originated from 14 GTG banded metaphases selected to minimize overlapping. Average to poor, rather than good or excellent, banding was selected a s being more representative of the general category of metaphases which would be encountered by a n automated system. Landmarks for No. 3 were determined as the minimum information needed to distinguish No. 3 from Nos. 1or 2. Although length and centromeric index were used in prebanding days to differentiate among the A group chromosomes, we were determining the minimal bands necessary for differentiation. We used the A groups to define the concept before testing it on groups more similar in length and centromere index, e.g., 7,8,9,X. For the karyotyping experiments, stylized chromosomes from the 12 metaphases were generated using the normalizing filter followed by heuristic algorithm preprocessing. This process was selected on the basis of the landmark experiment. The digital metaphase (original sampled chromosome images) and the stylized metaphase were both printed for evaluation. Each metaphase was assigned a random code and every stylized chromosome image was assigned a random number for the digital and stylized metaphases. Images were cut apart as they are in routine karyotyping. Each set of metaphase images was put in a separate envelope. Five cytogenetic technicians who each had at least one year’s experience, were asked to karyotype the contents of each envelope in their usual manner on standard forms. At most, these individuals had only casual previous exposure to stylized images. We used intact metaphases instead of isolated chromosomes to simulate actual practice. The only exception to routine cytogenetic procedures was the instruction to label a chromosome as “unidentifiable” if it really did not resemble any chromosome, rather than fit it into any empty slot by the process of elimination, as is common in routine karyotyping. “Unidentifiables” were counted as errors. We wished to remove bias due to lucky guesses if chromosomes were forced into empty karyotype positions. A total of 2,645 chromosome images were classified by each of the five technicians. For a given metaphase, if a cytogeneticist missed a certain number of the digital images, and had less than that number of similar errors on the stylized, the corresponding stylized images were considered not misclassified. Errors on stylized chromosomes are those not made on the corresponding digital images. Coinci-
43
STYLIZED CHROMOSOME IMAGES
METAPHASE (REAL IMAGE)
> SINGLE CHROMOSOME (REAL IMAGE)
(DIGITAL IMAGE)
DIGITAL IMAGE
I
1 DENSITY AND WIDTH
PROFILES
COMPUTER CLASSIFICATION (USUALLY CLASSIFY A L L PROFILES FROM THE METAPHASE TOGETHER) CHROMOSOME IMAGE
FIG.2. A generalized block diagram of the semi-automated karyotyping system to produce stylized chromosome images. Classification of profiles is usually done with all the metaphase data from each cell present, similar to methods employed in manual karyotyping.
dent errors are defined in the same way a s Lundsteen et al. (4-6); that is, made by a t least two cytogeneticists. This type of error doesn’t increase if another person misclassifies the same image. Isolated errors are those made only by one investigator.
Use of Stylized Chromosomes in Analysis of Chromosomal Aberrations Stylized images were prepared from a pair of No. 13 chromosome, one of which is deleted (20). Generation of stylized images has been incorporated into the Magiscan system (Joyce Loebl, Ltd., Newcastle-on-Tyne, England). Chromosomes from a n infant with del(4p) and from a mother heterozygous for a pericentric inversion in No. 5 (12), were transformed to stylized chromosomes on the Magiscan. For the del(4p), breakpoints and the amount of material deleted were to be determined. The diagnostic issue in the prenatal diagnosis of the fetus whose mother had a balanced pericentric inversion of No. 5 [46,XX,inv(5)(pl3q33)]was whether the fetus had inherited the balanced inversion or a n unbalanced recombinant, dup(q). Banding patterns would be similar in those two outcomes. Facilities The imaging analysis was done in the BioImaging Laboratory, in the Biomedical Engineering Department a t Case Western Reserve University. The images were sampled with a n Optronics drum
type optical scanner, processed by a PDP 11/34 computer, and displayed by a Gould-DeAnza system. Cytogenetic collaboration was provided at both Northwestern University and a t Case Western Reserve University Medical Schools.
RESULTS Effects of Preprocessing Algorithms on Stylized Chromosomes The original images, digital images, and the stylized chromosomes illustrating effects of preprocessing algorithms applied to density profiles, are shown in Fig. 3 using Nos. 1 and 7 a s examples. Results of grayscale remapping improve the apparent separation of dark and medium gray levels (Fig. 3D.11, while compressing the lighter regions of the grayscale (see the short arm of l p where the distal region loses resolution while conversely, the region above the centromere resolves into distinct bands). As mentioned in the Materials and Methods, a n infinite number of remappings are possible. The one shown here performed well in general on poor-quality metaphases to resolve dark bands and is a single example of a class of remappings that a n operator can easily design in a n interactive mode with the image processing system. We found grayscale remapping to be variable in performance: remapping functions that worked well with good-quality metaphases caused decreased band definition on poor quality material. This suggests that a n adaptive approach, choos-
44
MARTIN ET AL.
EVOLUTION OF AN IMAGE: A PAIR OF NO. I AND A PAIR OF NO. 7 CHROMOSOMES NO. I
NO. 7
A. Original, Light Microscope Image
6. Digital Image
C. Basic Stylized Image: Density Profile
Corresponding
D. Stylized Images after Preprocessing I) Grayscale Remapping
2) An Heuristic Algorithm
3) An Arbitrary Linear Filter (Highpass)
4) A Normalizing Filter
5) Normalizing Filter Followed by an Heuristic Algorithm
FIG.3. The appearances of the original (A),digital (B),and stylized (C,D) chromosome images of a pair ofNo. 1and No. 7 homologues from the same metaphase. Effects of various density profiles preprocessing algorithms are illustrated in D1-5.
S T Y L I Z E D CHROMOSOME IMAGES
ing the best function based on the quality of banding, should be implemented. This method also presented the best interactive capabilities of those investigated. An operator could easily adjust the remapping function for each image (similar in concept to adjusting “brightness” and “contrast” controls of a TV set). Results of a heuristic algorithm that was designed to maximize banding contrast within the body of the chromosome by emphasizing high frequencies are shown in Figure 3D.2 The heuristic algorithm improved some band definition; e.g., the distal G-positive band in l q and homologue similarity improved for most profiles. Differences from preprocessing by grayscale remapping were not startling. However, this algorithm was sensitive to input quality; incident banding patterns below a certain quality resulted in the introduction of artifact. Application of one form of linear filter, a n arbitrary highpass filter, is shown. Although the bands become very distinct with this procedure, a n arbitrary high pass filter is not suitable for preprocessing for classification because the simple contrast enhancement produces excessive artifact. This type of filter should be used for aberration detection as a final interactive step in a n editing mode. To remove the risk that artifacts will be mistaken for true aberrations, a standard, that is, a small or subtle known aberration might be used to adjust the filter. Then new aberrations may be sought in specimens from other cases. Because the frequency content of the incident metaphase will change due to varying quality, no one filter can be expected to perform well on all profiles. Thus, the “normalization” filter (Fig. 3D.4), unlike the arbitrary high pass filter, can adapt to varying quality of incident images. This operation obtains a measure of the banding quality of the incident metaphase and automatically designs a filter which will produce filtered profiles that approach that of a given “reference model.” Landmark bands were emphasized, and bands not used for classification were deemphasized. The results of the normalization filter were consistent: 1)a modest but definite improvement on “landmark” band contrast; 2) suppression of high-frequency bands, which can be considered “noise” to classification algorithms; and 3) little if any introduction of artifact. The controlled improvement in band contrast of the normalizing filter led to a n experiment where the profiles were first processed with the normalizing filter and then the heuristic algorithm. For the karyotyping experiment, the prior application of the filter had the beneficial effect of improving the “borderline quality” profiles such that the heuristic algorithm performed well on them, where it had not with the heuristic algorithm alone. Notice that chromosomes in Figure 3D.5 have a uniform width profile to determine if information on chromosome shape aids classification. Cytogeneticists comparing stylized chromosomes without vs. with width variation said they could classify those exhibiting width variation more readily.
45
Use of Stylized Chromosomes For Karyotyping Tables 1 and 2 summarize the stylized karyotyping results (the detailed results of the digital image karyotyping, 0.6% errors are presented elsewhere 3). Table 1 is the standard confusion matrix showing the distribution of the errors made for each chromosome and by each technician. Table 2 shows 1) the multiple error breakdown and 2) misclassification by chromosome group. The multiple error breakdown shows which chromosome types were misclassified just once, or by one or more investigators. The number of misclassifications is the number of images that were not recognized while the number of errors reflects the total number of karyotyping mistakes. The total error rate for stylized chromosomes was 7.4% (19712,645). The investigator error range was 5.5%-9.8% However, as shown in Table 2, sixty-three errors were isolated errors; i.e., chromosomes that were misclassified by just one investigator. Isolated errors, particularly on digital images, seemed to have been associated with carelessness. This explanation is proposed because the information content of those chromosomes was confirmed by the other four participants and because those technician recognized the same chromosomes in review after the experiment. Removing isolated errors from the total gives a 5.1%(1197-631/2645) error rate. There was variation among errors contributed by the 12 metaphases (0.5%-18.3%). The confusion matrix in Table 2 indicates that approximately 47% of the errors originated for the C + X group chromosomes, whereas only 34% would have been expected based on their proportion in the data set. The errors were discussed with the cytogeneticists who reviewed the nature of their errors; several explanations were found for most of the misclassification. Errors in the profiles caused by curved chromosomes. Computational error in the profiles caused by the failure to follow the chromosome contours accurately can best be described as producing “glitches” in the waveforms. These are especially confusing to the cytogeneticists in the width profiles where the glitches are more likely to be seen. Once aware of the causes of these glitches in the shape of the stylized images, the cytogeneticists thought that they would be able to compensate by relying more on the banding pattern than the overall appearance of those images that “looked strange.” However, i t suggests that automated karyotyping may benefit from further correction of curvature. Misclassification resulting from segmentation errors. Many of the misclassifications of the smaller chromosomes (E-G,) were due to the fact that the relative sizes of the chromosomes were in error due to the part of the image processing that is responsible for differentiating the boundaries of the individual chromosomes from the background (thresholding). For example, eight No. 19 chromosomes were classified as 16’s (Table 1) because the relative sizes were not main-
46
MARTIN ET AL. Table 1 CEassification of Stylized Vs. Digital Images From the Same Real Chromosomes"
Digital 1
2 3 4
1
3
2
5
4
6
7
8
9
Stylized 10 11 12 X 13 14
15
16
17
18 19 20
21
Total 5% errors 22 Y Unib stylized errors
107 3 5 105
3 5 0 6 3
110 1
5 2 112
114
5 6 7 8 9
1 109
1
10 11 12
1
1
1
4
5 115
1
1
4
3
2 113 2 1 119 76
5
13 14
113
1
3 2 4
3 5 105
2 3
18 19 20 21 22
3
111
8 1
101 2 5 107
1
1
1
1
1 1 1
1
1 107 1
Y
4
4
1
3
5
7.4
11
1
109
1
17
5
197
2
1
104 8 9 108
16
1142430
2
2
0
16 23 29 2 2 1 14
15
Uni Total errors digital
6
1 104 6 10 1 92 15 9 15 86
X
2.7 4.5 0 5.0 2.6 0.9 13.3 20.0 25.2 1.7 1.7 0.8 15.6 1.7 9.6 10.0 5.2 7.5 8.7 12.2 7.0 2.7 9.6 30.0
9
1 1 4
8
10
6
II
10
2 1
2 5 104 1 1 14
2
6
8
6
14
3
12 6 9 10 14 8 3 11
"Cytogeneticists 'Uni = unidentifiable.
tained. Improvement might be achieved if the cytogeneticists, not the engineers, selected the threshold. Confusion and artifact caused by enhancement. The contrast enhancement actually caused confusion in some cases. The cytogeneticists were not accustomed to seeing some low contrast bands or were misled by good contrast between bands that are usually hard to distinguish in normal photomicrographs. When confronted with these errors, the cytogeneticists agreed that once aware of the presence of these bands, they could adjust their decision-making. As stated above, the heuristic algorithm used to preprocess the profiles introduced artifact in poor-quality profiles. This caused most of the errors between classes 7-9, and X. Discussion with cytogeneticists brought out karyotyping features for these classes which had not been incorporated into the algorithms. Some of the participants relied on the relative positions and intensity of the distal dark band of the short arms; i.e., its more median position in 9p than 8p. An improved method of first applying the normalizing filter was later developed based on this information.
Analysis of Chromosomal Aberrations For determination of breakpoints and the amount of material deleted in the 13q, stylized profiles were adequate (10,181 (Fig. 4A,B). However homologue dissimilarity was unacceptable. Stylized chromosomes produced on the Magiscan IIA are shown in Figure 5. Similar stylized chromosomes were prepared from 1)blood cultures of a n infant with a del(4p); and 2) blood and amniotic fluid cultures from a mother with a inverted No. 5 (12). For the deletion, crude analysis of breakpoints and the amount deleted could be done (Fig. 6). The diagnostic issue in the case of inverted No. 5 was whether the fetus had inherited the balanced maternal inversion or a n unbalanced recombinant. In Fig. 7B a pair of chromosomes from the maternal chromosomes are shown. The density profile, stylized chromosomes, and digitized image from the normal 5 are shown on the left; the rearranged chromosome, on the right. Although the rearranged chromosome in both the mother and fetus could clearly be distinguished from the normal 5 by use of stylized im-
47
STYLIZED CHROMOSOME IMAGES
Table 2 Isolated and Multiple Error Breakdown for the Classification of the Stylized Chromosomes
Chromosome no. 1 2 3 4 5 6 7 8 9 10 11 12
X 13 14 15 16 17 18 19 20 21 22 Y Total
1 1 3 2 1 1 5 3 6 2 2 1 4 2 3 5 4 4 2 1 6 3 1 1 63
No. of times each chromosome misclassified cytogeneticist 2 3 4 1
5
1
2 1 1 1 3
3 6 3
1
3
1 2 1 1 4 4 1
2 1
2
1 1
1
1
26
20
2
1 1 3
No. of images misclassified 2 4 0 4 2 1 9 10 14 2 2 1 8 2 6 8 5 6 6 6 7 3 4 2 114
ages, comparisons of the maternal inversion (Fig. 7A) to the fetal rearranged 5 was inconclusive. The fetus was determined to be unbalanced by other means (12). The problems in the Magiscan stylized images appear to be poor normalization of bent chromosomes and inconsistent thresholding a t the ends.
Total no. of errors 3 5 0 6 3 1 16 23 29 2 2 1 14 2 11 12 6 9 10 14
8 3 11 6 197
Misclassified by group A 6 (5.3%)
Average error: total chromosome classified per group, % 1.8
B 6 (5.3%)
2.7
C 47 (41.2%)
5.9
D 16 (14.0%)
4.7
E 17 (14.9%)
5.0
F 13 (11.4%)
5.7
G 9 (7.9%)
4.0
114
cause difficulties in computer aided karyotyping. Using stylized chromosomes, we found that primary and secondary constrictions, width variations, presence or absence of certain key bands, and relative staining intensity of certain bands were important for karyotyping accuracy. Human karyotyping using density profiles has been DISCUSSION assessed by Lundsteen and colleagues (4-6). They reIn this paper we utilized stylized chromosomes to port a classification error of 0.5% on density profiles evaluate improvements in chromosome classification with the centromere marked. However, they may have by automated karyotyping effected by preprocessing excluded more of the bent chromosomes than we did, the density profiles measured on real chromosomes. based on the number of profiles analyzed per The objective in preprocessing the profiles from a metaphase (average 41 for their data, 44 for ours). Furmetaphase of poor banding quality is to produce styl- thermore, we included a n “unidentifiable” class and ized chromosomes with enhanced banding contrast to scored i t as a n error, rather than assigning i t on the extract information difficult to see by the human visual basis of exclusion, as is routinely done, and as Lundsystem. Preprocessing elongated chromosomes with steen did. This was to avoid the bias of correct guesses. many bands has as its goal reduction of information to Another important difference between Lundsteen’s exlandmarks essential for classification (karyotyping). periments and ours is that he was the only karyotyper, Bands can subsequently be reintroduced to detect ab- and he had training on another profile data set. Our cytogeneticists had no previous experience with stylerrations. We tested the assumption that the width and the ized chromosomes and showed considerable variation density profiles contained all the information neces- in their karyotyping accuracy using stylized chromosary for karyotyping. That is, because the stylized im- somes. For these reasons, our results are not strictly ages exhibit this information in a manner easily inter- comparable to those of Lundsteen. We found stylized chromosomes useful, as a common preted by cytogeneticists, any difficulty in manual karyotyping relates directly to the information content language between biomedical engineers and cytogenetavailable to the computer, and would also be likely to icists in determining the minimum graylevel (16-32)
48
MARTIN ET AL.
FIG.4. A Pairs of No. 13 chromosomes (homologues) from three metaphases (GTG banding). The deleted 13 is shown to the right. B: A pair of stylized chromosome images from the same patient.
and sampling density (11-20 samplesipm) necessary for cytogeneticists to identify chromosomes (3,181. Evaluation of preprocessing techniques for chromosome band enhancement was facilitated, specific improvements were suggested by the appearance of the stylized images, and insight was provided into mechanisms of human chromosome identification and classification. This information can be incorporated into development of automated cytogenetic systems (7,13,16). For example, a n optimal goal for classification algorithms is to enhance landmarks and reduce noise, whereas, preprocessing algorithms for detection of chromosomal band aberrations must be directed toward delineation of the smallest band capable of being reproducibly resolved. Stylized chromosome images with distinct band boundaries and perimeters are capable of exhibiting band resolution varying from the minimum landmarks required for classification to the maximum distinguishable for aberration detection. Hypothesis testing will be more efficient using these than density profiles (8,111.
We generated stylized images from abnormal chromosomes I(del(13q);del(4p); inv(5)] to perform quantitative analysis and hypothesis testing. Limitations of this application of stylized images to analysis of many small bands include the necessity for high-quality original images and relative homogeneity of staining of homologues within a metaphase. Homologue dissimilarity persisted in the del(l3q) case, although breakpoints and area of material deleted could be determined. Similar analysis using stylized chromosomes of the del(4p) generated on the Magiscan was satisfactory, but conclusive matching of the rearranged chromosome No. 5 in mother and fetus could not be achieved. The algorithm in the Magiscan was affected by minor bends in chromosomes, and did not achieve reliable thresholding so that ends of homologues were reproducible among cells. Gray levels can be converted from a 0-255 scale to pseudocolors, which reflect differences in underlying genetic architecture within G band negative material which are not discernable to the human eye.
49
FIG.5 . A karyotype of stylized chromosome images from t h e Magiscan (Joyce Loebl, Ltd.1.
FIG.6. Stylized chromosome images of a pair of No. 4 chromosomes (homologues) Magiscan IJoyce Loebl, Ltd.1. The deleted 4p is shown to the right. Centromeres are marked by a white band. Black borders are a t termini of the chromosomes and do not represent chromosome bands.
If stylized images could be shown to represent accurately the underlying real chromosomes, that is, na significant information is lost, another use for them would be as hard copy in automated systems. Cytoge-
neticists may even prefer karyotypes made from stylized chromosomes because of their normalized appearance. These would be useful for composite karyotypes, aggregates of chromosomes from different cells, which have been proposed for automated karyotyping rather than correcting for overlapping chromosomes (9). (Composite karyotypes are routinely employed by cytogeneticists to analyze aberrations.) Stylized chromosomes should prove useful for analysis of prophase chromosomes where the longer chromosomes bend and overlap, and the presence of large numbers of bands makes manual analysis tedious, inaccurate and subjective (15).Stylized chromosomes and composite karyotypes should provide a method of representation of prophase chromosomes which would facilitate comparison of band patterns after straightening and normalization. For this application, stylized chromosomes have the potential of extending the capabilities of the human.
50
M A n T I N ET AI,
U
5 FIG.7. A: Break points of a n inverted No. 5 (12). B: Density profiles, stylized chromosome images, and digital images, from a pair of No. 5 chromosomes processed by the Magiscan (Joyce Loebl, Ltd). The
inverted No. 5 is shown to the right. Its shortened p arm relative to the normal homologue, and the increased negative G band on the q arm, are reflected in the stylized chromosome.
IN MEMORIAM Our colleague Walter Johnson’s untimely death caused a great loss to modern cytogenetics and to his friends and colleagues.
JL Synthetic chromosome images a s aids to semi-automated cytogenetic analysis. Am J Hum Genet 31:104A, 1979. 11. Martin AO, Miller L, Simpson J , Thomas CW, Rzesyotarski, MS, Elias S, Sarto GE, Patel VA: Localization of the nuclear organized by computer-aided analyses of a variant No. 21 in a human isolate Hum Genet 48:211-219, 1979. 12. Martin AO, Northrup H, Ledbetter DH, Trask B, van Den Engh G, LeBeau MM, Beaudet AL, Gray JW, Sekhon G, Krassifkoff N, Booth C: Prenatal delection of 46,XY,rec(5),dup qjnv(5Kp13q33) using DNA analysis, flow cytometry, and in situ hybridization to supplement classical cytogenetic anaIysis. Am J Med Genet 31: 643-654, 1988. 13. Mendelsohn ML (Ed): Proceedings of the Workshop on Automation of‘ Cytogenetics, Asilomar, California. Lawrence Livermore Laboratory Technical Report CONF-751158, 1075. 14. Osterlinck A, VanDaele J, DeBoer J , Dom F, Reynaerts A, Van Den Berghey: Computer aided karyotyping with human interaction. J Histochem Cytochem 25:754-762, 1977. 15. Piper J: Interactive image enhancement and analysis of prometaphase chromosomes and their band patterns. Anal Quant Cytol Histol 4:233-240, 1982. 16. Piper J, Lundsteen C: Chromosome analysis by machine. Trends Genet 3:309-313, 1987. 17. Rzeszotarski, MS: Unpublished technical report on chromosome image analysis. Department of Biomedical Engineering, Case Western Reserve University, Cleveland Ohio, 1981. 18. Rzeszotarski, MS, Thomas CW, Martin AO: Sampling consideration in human chromosome images. Proceedings of the San Diego Biomedical Symposium, Martin J I (ed). 1978, Vol 17, pp 413-
LITERATURE CITED 1. DeBoor C: A Practical Guide to Splines. Springer-Verlag, 1978. 2. Gonzalez RC, Wintz P: Digital Image Processing, Chapter 4. Ad-
dison-Wesley, Redding, MA, 1977. 3. Isenstein BS: Stylized chromosomes: A chromosome imaging technique with application to automated cytogenetics. Master’s Thesis, Case Western Reserve University, Cleveland, Ohio, 1981. 4. Lundsteen C: Aspects of automated chromosome analysis: Different representation of banded human chromosomes and their cytogenetic evaluation. Ph.D. thesis, Copenhagen, 1979. 5. Lundsteen C, Granum E: Visual classification of banded human chromosomes; 11. Classification and karyotyping of integrated density profiles. Ann Hum Genet 40:421-442, 1977. 6. Lundsteen C, Granum E: Visual classification of banded human chromosome 111. Classification and karyotyping of density profiles described by band transition sequences. Clin Genet 15:430439, 1979. 7. Lundsteen C, Martin AO: On the selection of systems for automated cytogenetic analysis. Am J Med Genet 32:72-80, 1989. 8. Lundsteen C, Gerdes T, Philip K: A model for selection of attributes for automatic pattern recognition. Stepwise data compression monitored by visual classification pattern recognition. 15:243-251, 1982. 9. Lundsteen C, Gertes T, Maahr J : Cytogenetic analysis by automatic multiple cell karyotyping. In: Automation of Cytogenetics: Advances in Systems and Technology, Lundsteen C, Piper J (eds). Springer Verlag, New York, 1989. 10. Martin AO, Thomas CW, Rzeszotarski MS, Isenstein BS, Simpson
417. 19. Sprawl P: Physical Principles of Medical Imaging. Aspen Publishers, Rockville, Maryland, 1988. 20. Wilson L, Hodes B, Martin AO, Simpson J L , Ogata E: Cytogenetic analysis of a case of 13q-syndrome. J Pediatr Ophthalmol Strabismus 17:63-67, 1980.