clonal antibodies has allowed the development of sensi- tive assays for such ... A complete set of software ... Company, Needham Hts, MA), at 200g for 10 min (21,221. The CCB were ..... cence microscope and image analysis software that in-.
0 1989 Alan R. Liss, Inc.
Cytometry 10:256-262 (1989)
A Digital Image Microscopy System for Rare-Event Detection Using Fluorescent Benjamin R. Lee, David B. Haseman, and C. Patrick Reynolds3 Department of Pediatrics and the Jonsson Comprehensive Cancer Center, UCLA School of Medicine, Los Angeles, California 90024 (B.R.L.,C.P.R.);Department of Radiology, Dartmouth Medical School, Hanover, New Hampshire 03756 (D.B.H.) Received for publication April 18, 1988; accepted June 11, 1988
Instrumentation for rare-event analysis should be capable of reliably detecting infrequent cells (< 1:10,000) while both excluding false-positive signals and including true positive cells found in multicell clumps. We have developed a digital image microscopy (DIM) system in which a cytospin of 2 million cells is scanned with an intensified video camera (ISIT) using an IBM PC AT microcomputercontrolled microscope stage. PASCAL software controls the stage and analyzes video input, storing the location of positive cells to magnetic disk. The user can then “replay” each positive cell under computer control for either visual confirmation or analysis using other fluorescent probes. The computer re-
Detection of infrequent (< 1:10,000)cells in mixed cell populations is necessary for monitoring “minimal residual disease” in leukemia and lymphoma patients (1,8,10,12,17,18,20,23,25,30,36,40), detecting small numbers of metastatic cells in solid tumor patients (3,6,9,16,27,29,34), evaluating the efficacy of purging procedures designed to remove malignant or alloreactive cells from bone marrow harvested for transplantation (31-33,371, identifying fetal cells in maternal circulation (4,5,11,13), and assessing the frequency of mutational events (39,411. The high specificity of monoclonal antibodies has allowed the development of sensitive assays for such “rare-event analysis” using flow cytometry or manual microscopy, both of which are capable of detecting as few as one target cell per 100,000 background cells (27,361. In addition, a computer-controlled microscope system has been described that can detect mutant red cells a t a frequency of 1 in 10 million cells (41). The sensitivity of such assays is limited by several factors, the two most important being the number of cells analyzed for each specimen and the frequency of false-positive or false-negative signals generated by the assay. We wished to develop a n instrument that would facilitate such rare-event analysis, especially in bone mar-
quires 24 min to scan a cytoprep of 2 million cells, while playback for visual confirmation by the user averages 5 min. Using Hoechst33342 premarked cells seeded into bone marrow as a model system, we found that the DIM system reliably detects one target cell per million marrow cells. With appropriate immunological markers, this system will aid in evaluating bone marrow purged of tumor cells prior to transplantation and should also be useful for detection of minimal residual disease in blood or bone marrow from patients with leukemia or solid tumors. Key terms: Microcomputers, image analysis, microscopy
row. Such a system would need to analyze a large enough number of cells to provide the desired sensitivity, yet perform the assay in a short enough period of time to make the system practical. In addition, to ensure accuracy, the system would need to confirm positive signals as true events. We describe here a system that uses a microcomputercontrolled scanning stage on a fluorescence microscope to scan cells prepared by cytocentrifugation. Intensified video signals of fluorescent cells from a microscope are detected by a microcomputer using digital image analysis, and then confirmed visually by the user to permit semiautomatic detection of infrequent cells.
‘This investigation was supported in part by Naval Medical Research and Development Command Work Unit MF58.527.0004, NCI grant CA12800, and a grant from the California Institute for Cancer Research. LCDR Reynolds is assigned to the UCLA School of Medicine from The Department of the Navy. ‘The views expressed in this article are those of the authors and do not reflect the official policy or position of the Department of the Navy, Department of Defense, or the U S . Government. ‘Address reprint requests to Dr. C. Patrick Reynolds, Division of HematologyOncology, Childrens Hospital of Los Angeles, 4650 Sunset Blvd., Los Angeles, CA 90027.
FLUORESCENCE DIGITAL IMAGE MICROSCOPY
MATERIALS AND METHODS Hardware and Software The system is based on a n IBM PC AT microcomputer (6 Mhz 80286) equipped with twin 10 MB Bernoulli cartridge disk drives (Iomega, Roy, UT), formatted with a n interleaf factor of 8, thus increasing the speed of data storage. Software was developed in Turbo PASCAL (Borland International, Scotts Valley, CA) using serial input/ output (UO) procedures from the Turbo Asynch Tools Package (Blaise Computing Inc., Berkeley, CA) to control the scanning stage MDACE 1000 motorized stage controller (Ludl Electronics Incorporated, Hawthorne, NY) via the serial interface. A complete set of software tools to control the scanning stage, scan various patterns on a slide, and record fields of interest has been described elsewhere (19). Detection capability of fluorescent target cells is provided by a Dage MTI ISIT 66 intensified video camera (Dage-MTI Incorporated, Michigan City, IN) (35,43) on a Leitz Orthoplan Fluorescence Microscope (E Leitz, Rockleigh, N a n .The KS-330 video signal of the ISIT Camera is digitized by a PCVISION real-time digitizer and display board (Imaging Technology Incorporated, Woburn, MA) located in the IBM PC AT. The PCVISION board digitized a n image of 512 x 480 pixels (a 2 pixellpm resolution using a ~ 2 objective 5 lens) with 256 gray levels at a rate of 30 frame& (7,28). The aspect ratio of the 512 x 480 pixel image matches that of a standard video image. Hardcopy of scan results was generated at a resolution of 300 dotslinch by a LaserJet plus printer (Hewlett Packard, Boise, ID). Preparation of Cells for Analysis Cell preparations consisted of Hoechst 33342 (H342) premarked leukemia cells (MOLT-3, from American Type Culture Collection) (15) seeded into human vertebral body marrow as previously described (32). To provide optimal detection of the H342-marked cells, the microscope was equipped with a 100 W mercury lamp and a UV excitation (350 nm)/blue emission (460 nm) “D” cube. Various coatings for glass slides were investigated for the best cell attachment, including wet poly-Llysine (26) and dry poly-L-lysine (14,42) coated slides, albumin coated slides (22), and silanated slides (24). Dry poly-L-lysine coated slides were found to give the best cell-to-slide attachment. Cytospins were prepared by centrifuging marrow-leukemia mixtures onto poly-L-lysine coated microscope slides, prepared by covering one surface with > 100,000 MW poly-L-lysine (Sigma Chemical Co., St. Louis, MO) (lmg/ml in 0.1 M potassiumphosphate buffer, pH 7.0) and allowing the solution to dry (42). Cells were centrifuged using modified Centrifugal Cytology Buckets (CCB, International Equipment Company, Needham Hts, MA), a t 200g for 10 min (21,221. The CCB were modified by placing a silicon rubber gasket between the slide and the cell chamber, defining wells for three cell preparation areas of 13 x 15 mm (Fig. 1).Marrow leukemia mixtures were adjusted so that the appropriate number of cells for each well (usu-
257
B
FIG.1. Centrifugal Cytology Buckets (CCB).The CCB were modified by placing a silicon rubber gasket (B) between the slide (C) and cell chamber (A). A second silicon rubber pad (D) was placed beneath the slide as a cushion. Note that three cytoprep areas are defined by the CCB on each slide.
ally 2 million) was suspended in 0.4 ml of Iscoves DMEM (MA Bioproducts, Walkersville, MD) + 10% fetal calf serum, which was then pipetted into the well. The pipette was rinsed, and the rinse was added to the well. After centrifugation, medium was aspirated via the outlet port of the CCB and 70% methanol added to the inlet port. After 10 min, the bucket was removed from the centrifuge carrier, inverted, the slide carefully removed, and a coverslip placed over the cytoprep areas using a nonfluorescent mounting medium (Aquamount, Lerner Lab, New Haven, CT). RESULTS To provide a n efficient detection system, a method for placing a large number of cells on a slide for analysis was required. We investigated the effect of various coatings for glass slides to improve cell attachment during cytocentrifugation including wet poly-L-lysine, dry polyL-lysine, albumin, and silanated coated slides. We found that dry poly-L-lysine coated slides gave the best cell-toslide attachment for bone marrow-leukemia mixtures; using the centrifugal cytology buckets (CCB), we could efficiently deposit up to 2 million celldprep area. The
258
LEE ET AL. PCVision Frame Grabber
Intensified
PC AT Computer
F3
/ I
Fluorescence
Motorized
Stage Controller
Disk S t o r a g e
FIG.2. Configuration of the digital image microscopy (DIM) system for detecting infrequent cells. PASCAL software controls microscope stage movement and analyzes images from the intensified video camera. Stage coordinates of positive signals are saved to a disk file so that the user can re-examine (under computer control) each field containing a positive event.
CCB also provided a rectangular cytoprep area that was ideal for automated scanning. Careful observation of supernatant from the cytobuckets aspirated after centrifugation failed to reveal any loose cells. As a further test, H342-marked MOLT-3 cells were seeded into marrow at a concentration of 10 cells per million, and then 0.5, 1, 2, and 3 million total cells/prep were deposited using the cytobuckets. We found the deposition of leukemia “tracer” cells to be linear until greater than 2 million cells were deposited. Thus, if one limits each preparation to 2 million cells or less, the number of cells loaded into the cytobucket is effectively the number of cells analyzed. Also, cytospins of 2 million cells or less provided a good distribution of cells on the slide for analysis, without cell stacking or clumping. As the cells in such a preparation are in a nearly confluent monolayer, the user can detect cell losses large enough to cause significant variance in analysis by inspection. The configuration of the system is outlined in Figure 2. Under control of the microcomputer, the scanning stage moved a slide in increments of one field of vision. To ensure that all deposited cells were digitized and analyzed by the microcomputer, overlapping fields of visions were analyzed, and the edges of the cytoprep area were overscanned (Fig. 3). The computer digitized each area to be analyzed into a 512 x 480 pixel frame buffer in the PCVISION graphics memory. Because the fluorescent signal from a positive cell averaged 20 pixels in diameter with a ~ 2 objective, 5 it was not necessary to sample each of the 245,760 pixels in the frame buffer. Therefore, for images captured with a ~ 2 microscope 5 objective lens, a “sieve technique” was employed, which consisted of sampling every 8th column of every 15th row of the digitized image for positive cells. Use of the “sieve” algorithm reduced the amount of time required to analyze one frame buffer by 82%. To further speed analysis of the digitized image, values of one PCVISION input look-up table were subjected
I
I
C C B Slide FIG.3. Scanning pattern used by the system. The inset shows an enlargement of a partial scan area from a cytoprep. Circles represent the overlapping circular fields of vision of the microscope. Squares represent the area digitized by the PCVISION frame grabber into a 512 x 480 pixel image. A slight overlap of the square digitized fields (not shown) prevents missing cells located directly on the edge of the digitized fields. Overlapping circular fields of vision are scanned by the digitizer to ensure each portion of the cytoprep is analyzed by the computer.
Linear configuration of LUT before thresholding
///’
Ol/--.
0
values in memory
266
1
t I 0
LUT values a f t e r thresholding
I
I
I
I 100
1
1
I 966
Signal digitized f r o m c a m e r a
FIG.4. The configuration of the input look-up table (LUT) is shown before (inset) and after thresholding. Pixel values from 0 to 180 are assigned a value of 0, while pixel values from 181 to 255 are assigned a value of 2.
to a threshold algorithm (Fig. 4). The algorithm set all pixel values at or below the range of background fluorescence to 0, while bright fluorescent signals were set to a value of 2. By thresholding pixel values to 0 or 2, a positive event was easily determined by summing all pixel values tested and then checking the final total. Because events with a fluorescence level equal to background fluorescence were thresholded to 0, any positive events caused the final total to be greater than 0. If a positive cell was detected, the stage coordinates of the field with the positive event were then stored to a file on the cartridge disk drive. Calibration of the instru-
FLUORESCENCE DIGITAL IMAGE MICROSCOPY
259
4 I N I T W Z E INPUT
INITIALIZE OUTPUT LOOKUP TABLES
cII:IIII I DIGITIZE VIDEO SIGNAL INITIALIZE TOTAL PMEL
COUNT TO ZERO
1
I
ISELECT READBACK FILE
I
I
OPEN SECONDARY FILE TO SAVE “TRUE POSITIVE” CEIL LOCATIONS r
6
-READFROM(X,Y)DISKLOCATION FILE 1
1
1
I
HOVE STAGE TO
COORDINATES
INCREMENT INPUT MOK-UP TABLE MEMORY ADDRESS POSITVE CELL OR FAISE POSlTlVB CEIL?
J,
YES
U FIG.5. Flow chart of the procedure used to analyze a single digitized image for positive events. Note that every 8th pixel of every 15th row of the digitized image (after thresholding) is sampled for a pixel value >2 until the end of the frame buffer or until a positive “hit” is encountered. If a pixel value > O is detected, the computer stores the coordinates of the field for user confirmation of the event and then proceeds to the next field of vision.
ment and setting the threshold level is carried out using a sample with a large number of true-positive cells. The sensitivity of the system can be adjusted either by varying the gain on the video camera or through software by varying the threshold pixel values. Although not essential for analysis, the output lookup table was configured to display pixel values of 2 or greater at a value of 255, the highest pixel intensity, while values below the threshold were set to 0. This
provided a very high contrast image on the monitor, enabling the user .Go observe the scan in progress and to adjust the threshold gates using positive control samples. The program was capable of digitizing, thresholding, and analyzing one field of vision for positive events in 0.61 s. The algorithm used in analyzing a digitized image is outlined in Figure 5. After scanning the entire cytoprep, the computer “replays” each positive event for visual confirmation by the user (Fig. 6). The computer relocates the first positive event by reading stage coordinates from a disk file. The operator can observe the positive cell in the microscope
260
LEE ET AL. Positive Cells Detected 100 91
BO
60
40
20
FIG.7. Map of the results from a computerized Scan followed by user analysis of a single cytoprep. Each of the 1890 rectangles in the grid represents an area digitized and analyzed on the cytoprep. Dashes indicate false positives, while the larger polygons indicate confirmed positive cells. Note the outline of the foreign bodies that caused a large number of false positives in this cytoprep.
to confirm that it is a true positive or exclude it as a false positive. After the operator enters his decision on the keyboard, the computer moves the microscope stage to the next field of vision containing a positive event, allowing rapid user observation of each positive event. After the user codinns all positive cells, the computer generates a map of false-and true-positive events (Fig. 7) and displays counts of true- and false-positive events. Computer-controlled scanning of a single cytoprep (1890 digitized fields of vision) required 24 min, while semiautomatic playback for visual confirmation by the user averaged 5 min. Scanning with a lower power lens (i.e., ~ 1 0 did ) not significantly decrease the time to scan 2 million cells, as more pixels of the image had to be analyzed (because of the smaller pixel-to-cell ratio). Also, use of a lower power lens decreases fluorescence intensity (because of a lower numerical aperture) and provides a less suitable image for examination by the user when confirming positive events. To determine the reproducibility of the system in scanning slides, a cytoprep of 1million cells seeded with 100 H342-stained fluorescent cells was scanned repeatedly and the positive events confirmed by the user. As shown in Figure 8, the reproducibility of the system was quite good; we attribute the decrease in fluorescent cells that were detected in later scans to quenching of the H342 stain because of repeated analysis. The initial detection of < 100 positive cells did not represent false-negative results, but rather was due to