Experience with the Athena semi-automated karyotyping system

0 1990 Wiley-Liss, Inc. Cytometry 1159-72 (1990)

ExDerience With

the

Athena Semi-Automated

Karyotyping System'

Brian

H.

D.

L.

J.

Biomedical Sciences Division L-452, Lawrence Livermore National Laboratory, Livermore, California 94551

(B.H.M., J.D.T., M.L.C.); Department of Laboratory Medicine, University of California, San Francisco, California

94143 (B.H.M.); Faculty of Applied Physics, Delft University of Technology, Delft, The Netherlands (L.J.V., I.T.Y.)

Received for publication June 7, 1989; accepted September I, 1989

The traditional analysis and assembly of metaphase chromosomes into a karyo- gram is a slow and tedious process re- quiring intermediate photographic steps and manual manipulation of the chromo- some images. Much of this task is highly repetitive and readily lends itself to par- tial automation. Semi-automated karyo- typing systems now are being used in- creasingly in both clinical and research cytogenetic laboratories. Digital image processing techniques are used to cap- ture, manipulate, and make an initial classification of chromosome images. The Athena system uses commercially available components based on a Macin- tosh I1 personal computer. Digital image processing procedures automatically iso- late chromosome images from the meta-

phase and arrange them into a karyo- gram, using information about relative chromosome length, centromeric index, and banding pattern. The operator uses the intuitive graphics interface of the Mac- intosh computer to monitor each phase of the analysis, to resolve any problems in isolating chromosome images, and to rearrange the individual chromosome images while assembling the final karyo- gram. Athena is designed as a semi-auto- mated karyotyping system that is easy to learn and has sufficient power and ver- satility for routine use in the analysis of human metaphase chromosomes.

Key terms: Automated chromosome ana- lysis, karyotype, interactive editing, Mac- intosh computer interface

Numerous laboratories around the world are en- gaged in cytogenetic analyses of human cells for both clinical and experimental purposes. Historically, the karyotypic analysis of metaphase chromosomes has been a slow and tedious process. Metaphase spreads are located and photographed; film must then be developed and several prints made from each negative. Individual chromosome images are isolated by cutting the photo- graphs, identified by their morphology and banding pattern, and pasted onto a template to create the final karyogram. Cytogenetic personnel must be trained in cell culture and slide preparation, possess the patience required to scan numerous microscope slides, be com- petent in photomicrography and darkroom work, have the dexterity to cut and paste the images of the indi- vidual chromosomes to assemble the karyogram, and have the skill to recognize the individual chromosomes and their abnormalities.

Many of the procedures involved in producing a karyogram are amenable t o partial or full automation.

The amount of effort required may be significantly re- duced by automation. The single most important ad- vance is to use digital image processing techniques to avoid all photographic steps and to eliminate film de- veloping, enlargement, printing, and the manual cut- ting and pasting of individual chromosome images.

'This work was performed under the auspices of the University of

California Program for Analytical Cytology and the U.S. Department of Energy by the Lawrence Livermore National Laboratory under contract number W-7405-ENG-48, and with the partial support of the Commission of European Communities through the Medical and Health Research Program, Project Number II.I.lI13 and Amoco Tech- nology Company.

'Address reprint requests to Brian H. Mayall, Biomedical Sciences Division L-452, Lawrence Livermore National Laboratory, P.O. Box

60 MAYALI, ET AL.

These are time-consuming and tedious tasks that are especially suitable for automation.

Interest in applying automation t o analysis of hu- man chromosomes dates back three decades. Early sys- tems quantified the size and shape of individual chro- mosomes. Size is usually expressed as length o r area normalized to the other chromosomes in the same metaphase. Shape is measured by the centromeric in- dex, which is defined a s the size of the short or long arm relative to the size of the whole chromosome. Initially, the goal was only t o separate chromosomes into the major groups (6,151. When Caspersson discovered chro- mosomal banding patterns, he also appreciated the po- tential inherent in using these patterns for the auto- matic recognition of individual chromosomes (2,3). Major efforts were developed to classify chromosomes automatically using their banding patterns for identi- fication (4,51. Such early efforts demonstrated the fea- sibility of automation but were handicapped by the limited capabilities of available computers and digital image processing systems. Other groups used automa- tion to measure properties, such as DNA content, of chromosomes that already had been identified by their banding pattern (11). The early systems tended to be large, complicated, and difficult to use. While they played a key role in development of more advanced systems, they suffered from one or more problems that prevented their widespread use. Some of these prob- lems were cost, speed, ease of use, and limited ability to identify and characterize human chromosomes cor- rectly.

Recently, advances in computer science and technol- ogy, and in image analysis in particular, have made semi-automated analysis of metaphase chromosomes feasible, and several groups are working to develop truly functional systems. Such systems have been re- viewed comprehensively (9,10,13,19) and fall into two main groups. The first group consists of relatively ex- pensive integrated systems that frequently have spe- cial-purpose hardware. These provide a high degree of automation, including automatic location of metaphase spreads and creation of a final karyogram. Systems in the second group tend to be more modest in their price and capabilities and generally are adaptations to per- sonal computers of some of the features found on the more expensive systems. Examples of current systems include the Cytoscan (141, which is based on advances made by the MRC group in Edinburgh, Scotland, the Genetiscan (11, developed by Castleman and his col- leagues, and the Metachrome (7), developed by Ledley and Lubs. Lundsteen and colleagues have worked ex- tensively with the Magiscan system to adapt it to the requirements of computer-assisted karyotyping; more experience has been gained with the clinical applica- tion of this system than with any other system to date (8,12,18). Smaller systems based on personal comput- ers include Genevision (Applied Imaging, Santa Clara, CA) and Karyotec 100 (Amcor Electronics, Long Island City, NY).

Athena (16,171 is a semi-automatic system that be- longs to the second group in terms of cost and simplic- ity but is closer t o the first group in terms of its capa- bilities. The system uses standard commercially available components and is based on a Macintosh per- sonal computer (Apple Computer, Cupertino, CAI. The Macintosh computer has a n intuitive graphics inter- face that permits ready interaction with the image dis- play using the computer mouse device. The Athena program automatically isolates chromosome images from the metaphase. The operator edits the metaphase image to separate touching or overlapping chromo- somes, or to rejoin a chromosome that was fragmented. The program then arranges and displays the chromo- somes in a karyogram using information about relative chromosome length, centromeric index, and banding pattern. The operator readily rearranges the karyo- gram; possible errors in centromere location and chro- mosome orientation are corrected, and chromosomes are readily moved and exchanged.

In this paper, we report our experience using a de- velopment version of the Athena system in a research cytogenetic laboratory. We describe the important fea- tures of Athena and illustrate the steps used to create the karyogram. Finally, we discuss advantages and limitations of the system.

MATERIALS AND METHODS

Peripheral blood was drawn from a healthy male do- nor. Lymphocytes were stimulated with PHA, cultured for 52 hours, treated with Colcemid, harvested, and prepared for G-banding following standard procedures. The Athena system integrates software and hard- ware for the semi-automatic analysis of metaphase chromosomes. It is described in detail elsewhere (16,171. Briefly, the system uses a standard laboratory microscope equipped for transmission video micros- copy. Illumination is with a quartz halogen lamp using an unregulated power supply. Imaging is with a 100 x ,

1.30 NA, oil immersion plan apochromatic objective. A

heat filter eliminates infrared light, and contrast is optimized with a green band-pass filter. A small solid- state charge-coupled television camera (Model 4815, COHU, San Diego, CA) is attached to the trinocular head of the microscope with a “C” mount camera adap- tor. The overall magnification is such that a typical metaphase cell occupies about 70% of the diameter of the digitized video field. A small monochrome monitor displays the video signal and is used to assist in posi- tioning the cell and focusing its image.

The Athena workstation uses a Macintosh I1 series computer (Apple Computer, Cupertino, CAI, equipped with a n 8 bit video frame grabber and a 19 inch color monitor with 1,024 pixel by 768 line resolution. We use a 45 Mbyte removable disc system as a fast and ex- pandable medium for storing images and for exchang- ing them among different systems. Two versions of each image may be stored; one is the initial metaphase image prior to processing and analysis (240 kbyte), and

SE:MI-ALJTOMATE:I) KAEYOTYPING SYSTEM 61 \

I

I

([Continue]

I A

m i

a

a *

II

0

ll

0

0

@ Continue t o Edit Segmentation mask

0

0

It-.-

I=

(Resume)

[""I

I

I

I

I

I

FIG. 1. User interrupt display. This interrupt may be invoked at any time by clicking on the CANCEL button. In this example, the operator was editing a karyogram, part of which is seen in the back- ground. The operator wished t o return to Continue to Edit Segmen-

the other is the completed karyogram image that will be printed 170 kbyte). Athena images are printed by a

gray level thermal printer (LaserTechnics, Albuquer- que, NM). Output of lower quality can be printed by the standard Apple Laser Printer (Apple Computer, Cuper- tino, CAI, which is a PostScript (Adobe Systems, Moun- tain View, CA) device using dithering to simulate dif- ferent levels of gray.

The Athena software is a n integrated but highly modular program written in C. The program adheres closely to Apple developer guidelines and makes exten- sive use of the display and interactive features that are supplied in the Macintosh Toolbox.

SYSTEM OPERATION

Operation of Athena involves the Case Study Set Up phase and six functional phases: Capture Image; Seg- ment Image; Edit Segmented Image; Analyze and Karyotype Chromosomes; Edit Karyogram; and Print Karyogram. The program is organized as modules that correspond to these phases and are accessed in se- quence. The program control menu (see Figure 1) can be invoked from any functional phase and permits the operator to return to any preceding phase. The analysis then continues from whichever earlier phase is se- lected, including making a fresh start by capturing the image again, or even quitting the analysis completely. Each phase now will be considered separately.

Case S t u d y Set U p

Before karyotyping, the operator first defines the na- ture of the Case Study, such as peripheral blood, bone marrow, radiation damage, solid tumor, etc. Selection

tation mask. However, the operator also had the option to return to any of the other preceding stages or, by clicking on Resume, to return to the current Correct the Karyogram phase.

of a Case Study defines the user environment and the values of parameters that control the analysis. Exam- ples of such parameters include the method used to select the threshold; degree to which touching objects are separated automatically; size of objects to be elim- inated automatically as too big or too small; prototypic data base and assigned classification weights used to identify chromosomes from measurements of relative length, centromeric index, and position of bands; and specific comments about the Case Study that are t o be printed along with the karyogram. The operator has complete control of these parameters, which can be op- timized t o the special requirements of the current anal- ysis. For example, the default value for the elimination of small particles also removes double minute chromo- somes, which are not relevant in prenatal cytogenetics but are of importance in radiation and toxicity studies. Thus, in the latter studies, the value for elimination of small particles is set lower to prevent elimination of double minutes.

Settings used for particular studies are stored and can be re-used in subsequent studies with or without modification. The operator may select generic default settings, use settings previously optimized for a partic- ular Case Study, or may create a new environment by modifying the settings to meet the specific require- ments a t hand. In practice, we usually use the same Case Study setting, which, once optimized, rarely needs to be adjusted.

C a p t u r e I m a g e

The first function phase controls the frame grabber. The operator uses the monochrome monitor to assist in

62 MAYALI, ET AL.

centering and focusing a metaphase image. The system monitor has a parallel display of the output of the video frame grabber, mapped so that signal saturation is shown in red. Illumination intensity is adjusted until there is only the occasional red pixel. Clicking on a software button captures and digitizes a single frame from the camera using 8 bits (256 levels) of gray scale. Atheria can work with a 640 horizontal by 480 vertical image (North American standard) or a 768 horizontal by 512 vertical image (European standard). The user places a 512 horizontal rectangle anywhere within the bounds of the digitized image. Athena then stores and processes the image. Many metaphase images may be stored in succession for subsequent analysis on the same or another Athena system. Alternatively, each image may be analyzed immediately after it has been captured and while the metaphase spread is still visi- ble under the microscope. We find the latter mode to be preferable, particularly when analyzing difficult spreads, which the operator may wish to re-examine through the microscope.

Segment Image

The second functional phase uses automatic scene segmentation to isolate individual chromosome images from the metaphase. Athena uses a threshold t o create and display a binary image in which all pixels having a gray level darker than the threshold are considered to belong to objects (putative chromosomes), and all lighter pixels are considered to be background. A threshold is selected automatically by the program but may be readily modified interactively by the operator to optimize the separation between chromosomes and the background. Binary operators process the image to remove dirt and particulate debris, to separate touch- ing objects, to remove objects crossing the edge of the image, and to count and color the remaining objects. At this step, the screen appears as shown in Figure 2.

Edit Segmented Image

Following segmentation, Athena enters the interac- tive editing phase. The screen appears as is shown in Figure 3a. (The metaphase cell shown in this and the other figures is representative for this study and for the processing times reported below in Tables 1 and 2). The operator may now edit the segmentation mask using the mouse. To separate touching chromosomes, the op- erator clicks on the Cut button in the upper left corner,

moves the cursor to the appropriate place on the screen, clicks the mouse and moves the cursor to define a blue cut line, which stretches as the cursor is moved. An irregular boundary between chromosomes may be de- fined by clicking on a series of points to create multiple line segments. When the line is in the right place, the mouse is double-clicked, and the line of separation is complete (see Figure 3b). Similar manipulations allow the operator to Remove unwanted objects, Join frag-

mented chromosomes, or Duplicate overlapping chro-

mosomes. After duplication, the operator cuts one over-

lapping chromosome from the original pair and the second chromosome from the duplicate pair and then removes the redundant portions (Fig. 3c). If desired, the user may activate the Zoom feature of the display

controller to enlarge the region of interest. Finally, the options of Undo Last and Undo All enable the user to

recover from editing mistakes. At any time while edit- ing the image, the operator can click the Do It button

(the button now is labeled Recount) to have the sys-

tem update the object count and reassign colors in the right-hand image. Figure 3d shows the metaphase of Figure 3a after the operator has completed interactive editing.

Analyze and Karyotype Chromosomes

When the operator is satisfied with the edited image, the program proceeds to analyze and classify the chro- mosomes. First, the major axis is found for each chro- mosome, then the image is rotated so that the chromo- some is vertical. The centromere is located, and the chromosome is inverted if necessary. The length, cen- tromeric index, and banding pattern are measured, and these measurements are used by the classifier to identify the chromosome. Individual chromosomes are aligned on their centromeres and are placed on the karyotype board according to their initial identifica- tion as is shown in Figure 4.

Edit Karyogram

The user edits and corrects the automatic karyogram as required. To Move chromosome images, the opera-

tor clicks on a chromosome and uses the mouse to drag it to a new location. If the location is already occupied, the displaced chromosome is moved to a spare location a t the bottom of the karyogram or else, as a n alterna- tive option, the two chromosomes are exchanged with one another. In aneuploid cells with polysomy, the karyogram table expands automatically to accommo- date the extra chromosomes. Invert allows correction

of chromosome orientation, and Centromere allows

the operator to define precisely the position of the cen- tromere (Fig. 5). Rotate, Insert Text, and Place Tag

are not yet implemented.

Print Karyogram

When the operator is satisfied with the karyogram (Fig. 61, i t is prepared for printing by first clicking on the Print Setup button. Text is added to identify the

karyogram and the slide coordinates of the metaphase, to provide relevant clinical data, and to enter the cy- togenetic findings. The operator then selects the Print

option. The finished karyogram is written as a file, which may be either spooled for immediate printing as a background task while Athena processes the next cell or saved for subsequent batch printing. A full gray level printer produces near-photographic quality out- put. Alternatively, a less expensive standard laser printer gives serviceable but lower fidelity output be- cause i t has t o use dithering to emulate gray levels.

SEMI-AUTOXATED K A R Y O T Y P I N G SYSTEM

FIG. 2. Processing of the binary image to effect chromosome seg- mentation. Two images of the metaphase are seen. On the left is the original metaphase image displayed with automatic contrast stretch- ing. On the right is the binary image in which putative chromosomes, colored green, are being processed and segmented. At the upper left is

the control window in which the program reports on the progress of the analysis. The control buttons in this window allow the operator to

Cancel the analysis and return t o the interrupt menu shown in Fig- ure 1.

Figure 7 shows a karyogram and comments as printed by the two devices.

RESULTS

We use both subjective and objective criteria to eval- uate performance of the Athena system. Subjective cri- teria, which are notoriously difficult to quantify, refer to system characteristics such a s user acceptance, te- diousness, and fatigue. In our laboratory, we find that semi-automatic karyotyping with the Athena system rapidly becomes the preferred method. Only about one- quarter of the cytogeneticist’s time is spent waiting for the computer to complete computational steps (see Ta- ble 1). Cytogenetic technologists without any prior computer experience find that Athena is easy to learn and that its features are intuitive and natural. In large measure, this is because the program exploits features of the Macintosh user interface. An important benefit

is that Athena can be controlled entirely from the mouse. Thus, the system can be used in a darkened

room, and the operator has no need to look anywhere but a t the display. The keyboard is used only to create

a new Study and to enter optional comments on the karyogram. Additionally, the new operator quickly dis- covers how easy it is to recover from any mistakes made during the analysis, even including restarting the whole analysis.

The display has appropriate size and gray level res- olution for efficient karyotyping. The ease of moving chromosomes on the karyogram table is especially use- ful. Chromosomes may be given tentative assignments and then compared directly with normal homologs. This capability encourages testing of different assign- ments and gives operators increased confidence in their assignment of ambiguous or badly distorted chromo- somes.

FIG. 3. Interactive editing of the segmented image. a: The window

as seen immediately after automatic segmentation. Two images of the metaphase are displayed. In the larger display, the gray scale is op- timized by contrast stretching, and each individual chromosome or cluster of chromosomes is outlined in red. In the smaller image, a palette of colors differentiates the chromosomes and clusters. Chro- mosomes that are touching or overlapping are considered as one object

and are shown in the same color in the smaller image. The object count is given on the left. b: The window during the Cut option.

Multiple lines have been drawn t o Cut the many touching chromo-

somes. Note that the overlapping chromosomes in the lower left al- ready have been Duplicated. The interphase nucleus is displayed in

FIG. 3c: The window during the Remove option. The cut portions of

the duplicated overlapping Removal. Note that the blue Cut lines

from the previous screen have been replaced by red lines that show the boundaries of the separated chromosomes. d: The fully segmented

metaphase image after interactive editing. The program now finds 46 objects, and the Removed objects no longer are visible. The color map

in the upper right is redrawn to reflect the new object count. Clicking the Continue button causes Athena to move t o the karyogram phase.

66 MAYALL ET AL Table 1

T i m e to Complete Analysis Tasks (Excluding finding metaphase spreads)

performed Elapsed time"

Task by Mean Range

Acquire image

Capture Image Operator 0 5 0 0:36-1:26 Grab & Store Image Athena 0:02 0:02-0:02

image 0 5 2 0:38-1:28

Read & segment image Athena 0:37 0:31-0:49 Edit segmented image Operator 3:43 1:05-4:05 Analyze & karyotype Athena 0 5 3 0:35-1:06 Edit karyogram Operator 2:55 2:14-4:49 Print set up Operator 1:16 1:02-1:30 Print conversion Athena 0:42 0:38-0:44 Total elapsed time to acquire

Karyotype image

Total elapsed time to karyotype

image 8:47 7:25-10:08

Operator (user active time) 6:35 4:58-7:58 Athena (user inactive time) 2:12 1:49-2:30 'In minutes and seconds.

Table 2

Operator Editing Tasks

Task Mean Range

c u t 23 12-34 Duplicate 1 0-3 Remove 7 2-18 Join 0 0-1 Move 26 23-35 Centromere 16 9-27

Edit segmented image

Edit karyogram

Invert 8 3-16

Objective criteria, used to evaluate system charac- teristics, include speed, initial accuracy of segmenta- tion and classification, and the probable assignment of errors owing to deficiencies in either Athena or the material being analyzed.

Table 1 gives the time required to acquire the metaphase image and to complete each phase of its analysis. These times are the mean and range for anal- ysis of ten consecutive human metaphase cells located on a single slide of cultured male lymphocytes. Cells were selected because they showed adequate banding for manual analysis and appeared to be complete. No other selection criteria were used. The cells were ac- cessed into the analysis as they were located even though they usually had many touching or overlapping chromosomes. For one of the cells included in these times, the operator made a n error in editing the seg- mented image. This error was not detected until well into editing of the karyogram. The operator returned to correct the segmented image and then completed the correct karyogram. The added time for this corrective sequence is included in the data of Table 1 a s the upper limit of the ranges.

Table 1 also gives the total time taken from the ini-

tial command to load the metaphase image from disc storage until the command is repeated for the next cell. This total excludes the time required t o locate the metaphase, which varies greatly, and the time re- quired to adjust the microscope, capture the image, and store it on disc.

The data of Table 1 show that 75% of the analysis time is taken by the user-interactive phases, while only

25% is taken by computer processing. Table 2 summa- ries the editing actions that took most of the operator's time and that were required to complete the karyo- gram analysis. Almost one-half of the chromosomes touched one another after initial segmentation (number of counted objects ranged from 21 to 37, mean = 28) and so required interactive separation. On aver- age, each cell had one pair of overlapping chromosomes that had to be duplicated; four cuts and four removes then were needed to resolve the pair into its constitu- ent chromosomes. Occasional interphase nuclei, and very rarely dirt artifacts, also had to be removed. It is

unusual to have to join a chromosome except to correct an operator error, a s happened once in this series.

Most operator actions in editing the karyogram were moving chromosomes. Usually, these were to correct errors in initial classification, but sometimes they were to test ambiguous chromosomes in different positions. About one-half of the centromere movements were minimal adjustments primarily to improve the esthet- ics of the final karyogram.

DISCUSSION

We first examine each phase of the Athena analysis to determine how time is spent and where there may be room for improvement. We then discuss the overall fea- tures of the Athena approach in the context of com- puter-assisted systems for karyotyping.

Capture Image (Operator)

The major portion of this time is spent centering the metaphase spread, optimizing focus, and adjusting the light intensity. Only a system with automated stage movement and focus would increase the speed of the first two steps. At present, the microscope has a n un- regulated power supply, which results in significant fluctuations in light intensity from image to image. Using a stabilized supply would decrease or eliminate the need to re-calibrate the illumination with each im- age. Actual image capture takes only 30 ms and less than 1 s is needed to write the image to disc.

Segment Image (Athena)

Segmentation of the image involves multiple image processing operations. The threshold for the gray level image is set automatically. The binary image under- goes extensive processing to remove dirt and small ar- tifacts and t o separate touching objects. Then the bi- nary map is colored so that each object has a separate and unique identity. The image processing algorithms are optimized for speed and efficiency; thus it is un-

SENI-AUTOMATED KARYOTYF'ING SYSTEM 6 7

Fro. 4 . Display of the initial karyogram. Each chromosome image is individually contrast stretched to provide maximum enhancement of

the bands. This gives images that often are better than those seen directly through the microscope. Note that there are many errors,

including misclassified chromosomes, inverted chromosomes, and chromosomes with misplaced centromeres. The operations Move, In- vert, and Centromere permit ready correction of such errors.

likely that this phase could be speeded greatly without utilizing a much faster processor.

Edit Segmented Image (Operator)

This phase is highly variable, depending almost en- tirely on the complexity of the selected cell. The G band protocol tends to produce broad chromosomes that touch one another more frequently than do non-banded chromosomes. Still, improved image preprocessing may significantly reduce the number of touching chro-

mosomes in the segmented image. Possibly the biggest single improvement would be to have interactive ad- justment of the selected threshold, thereby permitting separation of the greatest number of chromosomes without leading t o fragmentation or loss. It may be possible to use a more sophisticated image processing algorithm to separate touching chromosomes; however, it is unlikely that this would increase the overall speed, and the benefits would be less than those from optimiz- ing the threshold in the first place. Duplication of over-

FIG. 5. Operations to correct a n error in centromere placement. a: The centromere for this chromosome 4 was incorrectly located in the middle of the long arm. b: The chromosome has been inverted. c: The centromere is now located by the operator. d: Finally, the chromosome is aligned correctly with the karyogram.

lapping chromosomes and subsequent removal of re- dundant material remains a time-consuming step and one that is not readily automated.

Analyze and Karyotype Chromosomes (Athena)

Automatic analysis and karyotyping requires only one second per chromosome image. It is possible that some algorithmic enhancements may speed this phase, but it is unlikely that significant gains will be achieved in the near future.

Edit Karyogram (Operator)

In this phase, the operator corrects mistakes made by the karyotyping algorithm. Badly bent chromosomes are almost invariably misclassified. The planned addi- tion of a n algorithm for Athena to straighten chromo- somes should reduce this error. A large number of er- rors are due to faulty location of the centromere; an improved algorithm for locating the centromere could improve this substantially. Many errors can be traced to poor segmentation, when, for example, the operator creates a n incorrect cut line t o separate two touching chromosomes. Touching and overlapping chromosomes always give measurements that will be suspect. Sepa- ration of touching chromosomes can create a spurious constriction that may be interpreted a s the centromere. Regions of chromosome overlap are abnormally dark and presently are interpreted by Athena as a dark band.

Some errors are not easily explained and probably reflect deficiencies in the classifier and in the data base we used. The present classification algorithm is rea- sonably sophisticated although there is room for im- provement. I t recognizes the pair-wise logic involved in attempting to classify any chromosome and uses max- imum likelihood estimates to optimize assignments. However, different weights may be assigned to each feature, and it may be that the classifier should be made more hierarchical in structure so that particular features become important only a t the relevant phases of the classification. There seems to be a tendency for a major error with one chromosome to cause further er- rors that cascade through the analysis. Finally, i t is probable that the data base values used by the classi- fier should be optimized for each laboratory to compen- sate for differences in preparation and staining.

Print Set Up (Operator)

The time taken to prepare the karyogram for print- ing is almost entirely due to the operator entering de- scriptive comments into the karyogram report. It is possible that some of these may be standardized and incorporated as default comments for a small saving in overall time.

Print Conversion (Athena)

The system requires time to convert and store the karyogram and to initiate printing. Our development system used a Laser Printer and so first converts the

SEMI-AUTOM.4TEI) I<AKYOTYI’IiY;G SYSTEM

FIG. 6. Completed karyogram. The operator is satisfied with the karyogram and now is ready to

Continue to t h e analysis of‘the next cell.

karyogram to a Postscript file, which is spooled and is then printed as a background task while the operator starts analysis of the next cell. Gray level printers do not require Postscript file conversion and may save some time overall.

OVERALL SYSTEM

Athena has many advantages as a system for com- puter-assisted cytogenetics. The first is that it is easy to learn. People with little or no experience with com- puters or the Macintosh operating environment quickly and easily learn to use Athena. After a n initial explanatory session, new users can work their way through a n analysis simply by following the prompts on the screen. Many of the steps are self-explanatory. If a mistake is made, Athena allows the user to return to a previous step and resume operations. Because this is accomplished quickly and easily, the user is able to explore alternate ways of editing the segmentation im- age. This ability to return to a previous step in the analysis is also helpful in correcting editing errors, a s

may occur when cutting apart overlapping chromo- somes or the small F and G group chromosomes. The ease with which errors can be corrected has a n impor- tant psychological benefit in that the operator feels in full control of the analysis.

Unlike many semi-automated karyotyping systems, there is no need for the user to identify each chromo- some in advance or to indicate manually the position of the centromere and telomeres. Athena assigns each chromosome a position on the karyotype board. This means that, with suitable material, Athena can pro- duce a final karyotype from the initial digitized image with only a minimum of human interaction. Problem chromosomes remain unassigned. For abnormal kary- otypes, the karyogram board adapts automatically to accommodate polysomies and polyploidy to a maximum of 127 chromosomes.

It is relevant to compare the overall speeds of Athena and other systems, even though such comparisons are difficult because of differences in material and in cri- teria used. Lundsteen and colleagues report their ex-

SEMI-AIJTOMATED KARYOTYPING SYSTEM 71

20

c

21 22 X Y

FIG. 7. A digitized metaphase spread (see previous figures) and its karyogram. a: The original metaphase spread, digitized to 256 gray

levels without contrast stretching and printed on the thermal gray

scale printer. b: The corresponding karyogram output, also printed on the thermal gray scale printer. c: The same karyogram, but printed on a Postscript laser printer using dithering to simulate gray levels.

perience using the Magiscan system in their clinical laboratory and find that approximately 7 minutes are required for the complete analysis of a metaphase spread (8). Lundsteen and Martin reviewed several sys-

tems; they reported that karyotyping generally takes about 7-10 minutes speed and that throughput seems to be largely independent of the system used (9). These times are comparable to those given in Table 1 for Ath- ena. As with the other systems, we also find that most of the time is spent in operator interaction. Thus, in- creasing computer speed will do little to enhance throughput. Decreasing the need for operator interac- tion should, however, have a significant effect on the throughput of Athena and other semi-automated sys- tems and may be realized by improving the sophistica- tion of segmentation, analysis, and classification algo- rithms.

Athena relies upon a data base of identifying param- eters, obtained from previous chromosome images, to classify each chromosome and place it appropriately in the karyogram. This data base is adaptive and has the potential to be a unique strength of Athena. Each lab-

oratory may personalize this data base for their partic- ular application, a task that is accomplished by adding appropriate chromosome images to the data base as they arise in the course of normal operations. As the data base grows to reflect the experience of a particular laboratory, the ability of the classifier to identify chro- mosomes correctly should improve. Although it is un- reasonable to expect Athena to identify every chromo- some, only one-half of the chromosomes are assigned correctly when using unselected material as in the ex- periment of Table 1.

A problem that Athena shares with all automated analysis systems is that the classifier is always depen- dent on the quality of the input data. Thus, no matter how sophisticated the classifier program is, there al- ways will be problems identifying chromosomes that are touching, overlapping, or badly distorted. Many classification errors result from misplacement of the centromere. An improved algorithm for centromere lo- cation is being developed and should improve this phase of the analysis. In the meantime, the sophisti- cated operator may define the centromere artificially

72 MAYALI ET AL.

by pinching the chromosome outline while editing the segmented image. Flagging all chromosomes that have had to be separated, particularly overlapping chromo- somes, should permit the classifier to avoid misinter- preting spurious bands, The creation of a more rigorous hierarchical and rule-based scheme for chromosome classification, designed to mimic the cytogeneticist more closely, may also improve the correctness of the initial karyogram assignments.

The small F and G group chromosomes always cause problems. Not infrequently, their images are so faint and poorly defined that the program is unable to find their major axis and their image needs to be rotated by 90" before being analyzed. G group chromosomes fre- quently associate with other chromosomes in the metaphase cell, yet they are so small that inaccuracies in separating them may cause substantial errors in their analysis. In principle, it may be possible to im- prove the automatic analysis of the smallest chromo- somes; in practice, it is probably more reasonable to rely on the skill of the cytogeneticist to correct errors when they occur.

Semi-automated analysis of human metaphase chro- mosomes has several potential advantages compared to manual methods. Perhaps the most significant en- hancement, a t least in terms of time savings, is that semi-automated systems perform data capture and storage directly, thereby eliminating the need for film development and printing and the associated dark- room. Images are captured directly into the karyotyp- ing system and can be analyzed immediately if desired. Image contrast is adjusted automatically, so there is no need for exposure bracketing or preparation of multiple prints a t different exposure settings a s is customary with photographic analysis. Automatic contrast stretching, particularly if performed on individual chromosomes as done by Athena, further enhances the image, and may make it appear even better than the image seen through the microscope.

Substantial time savings are realized by the in- creased productivity that accompanies elimination of the darkroom. Our estimates of throughput improve- ment range from 50% to 100% or more, but this varies with the type of analysis. Other advantages include relief from much of the tedium involved in traditional cytogenetic analyses. This includes working in the darkroom and cutting and pasting individual chromo- some images from photographic prints of the metaphase cells. Another potential advantage is a n in- crease in accuracy, which is aided by the image en- hancement and contrast stretching to optimize the ap- pearance of each individual chromosome and by the ease of testing ambiguous chromosomes against a num- ber of other chromosomes in the karyogram.

To summarize, the Athena system combines com- puter power with a truly user-friendly operating envi- ronment. The result is a karyotyping system that is easy to learn and use, that has sufficient speed for rou- tine analysis of human metaphase chromosomes and flexibility for analysis of highly abnormal karyotypes, and that can adapt to meet the specific requirements of disparate cytogenetic laboratories.

LITERATURE CITED 1. 2 . 3. 4. 5. 6. 7. 8. 9. 10. 11. 12 13 14 15 16 17 18 19

Berry R, McGavran L: Impact of computerized imaging system on

a clinical cytogenetics laboratory: Six months experience. Pre- sented a t 25th Annual Somatic Cell Genetics conference, New Mexico, 1987.

Caspersson T, Lomakka G, Moller A: Computerized chromosome identification by aid of the quinacrine mustard fluorescence tech- nique. Hereditas 67:102-110, 1971.

Caspersson T, Lomakka G, Zech L: The 24 fluorescent patterns of the human metaphase chromosomes-distinguishing character- istics and variability. Hereditas 67239-102, 1971.

Castleman KR, Melnyk J , Freiden HJ, Persinger GW, Wall RJ: Computer-assisted karyotyping. J Reprod Med 17:53-57, 1976. Granlund GH, Zack GW, Young IT, Eden M: A technique for multiple-cell chromosome karyotyping. J Histochem Cytochem 24:160-167, 1976.

Ledley RS: High speed automatic analysis of biomedical pictures. Science 146216-223, 1964.

Ledley RS, Lubs HA: Description of the Metachrome system. Sup- plement to Report to European Working Group on Automated Chromosome Analysis, Berlin, 1987.

Lundsteen C, Gerdes T, Maahr J, Philip J: Clinical performance of a system for semi-automated chromosome analysis. Am J Hum Genet 41~493-502, 1987.

Lundsteen C, Martin AO: On the selection of systems for auto- mated chromosome analysis. Am J Med Genet 3272-80, 1989. Lundsteen C, Piper J , Eds: Automation of Cytogenetics. Springer Verlag, Berlin, 1989.

Mayall BH, Carrano AV, Moore DH 11, Ashworth LK, Bennett DE, Mendelsohn ML: The DNA-based human karyotype. Cytom- etry 5:376-385, 1984.

Philip J , Lundsteen C: Semiautomated chromosome analysis. A clinical test. Clin Genet 27:140-146, 1985.

Piper J , Lundsteen C: Human chromosome analysis by machine. Trends Genet 3:309-313, 1987.

Rutovitz D: Automatic chromosome analysis. Pathologica [Suppll 75:210-242, 1983.

Rutovitz D, Cameron K, Farrow ASJ, Goldberg R, Green DK, Hilditch CJ: Instrumentation and organization for chromosome measurement and karyotype analysis. In: Human Population Cy- togenetics, Pfizer Medical Monographs, Vol5, Edinburgh Univer- sity Press, Edinburgh, 1970, pp 281-296.

van Vliet U, Young IT, ten Kate TK, Mayall BH, Groen FCA, Roos R: Athena: A Macintosh-based interactive karyotyping sys- tem. In: Automation of Cytogenetics, Lundsteen C, Piper J (eds). Springer Verlag, Berlin, 1989, pp 47-66.

van Vliet U, Young IT, Mayall BH : The Athena semi-automated karyotyping system. Cytometry 11:51-58, 1990.

Wulf HC, Philip J: Semi-automatic karyotyping facility-a clin- ical test. Hereditas 105:37-40, 1986.

Zeidler J P : Automated chromosome analysis. Nature 334:635, 1988.