Morphologic Changes in Rat Urothelial Cells During Carcinogenesis: II. Image Cytometry

(1)

Morphologic Changes in

Rat

Urothelial Cells During

Carcinogenesis:

II. Image Cytometry

l

Ian

T. Young, Martin Vanderlaan,

Lois

K r o m h o u t ,

Ronald Jensen, Abigail

Grover,

and Eileen

K i n g

Pattern Recognition Group, Department of Applied Physics, Delft University of Technology, 2600 GA Delft, The

Netherlands (I.T.Y.), Lawrence Livermore National Laboratory, Biomedical Sciences Division, University of California, Livermore, California 94550 (M.V., R.J., A.G.), and School of Medicine, University of California San Francisco,

San Francisco, California 94143 (L.K., E.K.)

Received for publication April 26, 1983; accepted March 9, 1984

Improved early detection of neoplasia by screening of urothelial cells requires an understanding of the features distinguishing normal and neoplastic cell populations. We have begun a program of study based upon a rat model system for the controlled observation of early-stage lesions produced by the carcinogen N-butyl-N-(4-hydroxybutyl)- ni- trosamine. Cells dissociated directly from normal and malignant urothelium were characterized by conventional cytopathology techniques and by quantitative microscopy (for nuclear texture and nuclear and cyto-

plasmic size, shape, and stain content) to de- rive a comprehensive picture of bladder tumor development. By following the changes that occur in the dissociated urothelial cells we have found that the nuclear area, total nuclear stain, nuclear shape, and the nuclear chromatin change significantly over a 48-wk interval as the lesions progress toward malignancy.

Key terms: Image analysis, cytopathology, rat bladder carcinogenesis

Our goal is to quantitate urothelial cell morphology for the improved early detection of bladder neoplasia (1). The screening task is a potential target for eventual automation, but first, a thorough understanding of the characteristics of normal and neoplastic cell populations is required. We have used a rat model system for t h e controlled observation of early-stage lesions produced by a known carcinogen. In this study, cells dissociated directly from normal and malignant urothelium a n d stained with Papanicolaou stain were characterized by conventional cytopathologic techniques a n d by quantitative microscopy to develop a comprehensive picture of the urinary cytology. The pathology of the tumors, methods for preparing dissociated cell samples, and their evaluation by conventional cytopathologic criteria is described in the preceding paper (6).

MATERIALS AND METHODS Samples

Male Fisher 344 rats were exposed to the carcinogen

N-butyl-N-4(-hydroxybutyl)-nitrosamine (BBN), as de-

scribed in the preceding paper (6). Beginning at 7-8 wk

of age the animals received 0.05% BBN i n their drink-

ing water for 10 wk and were then returned to control

water for the duration of t h e experiment. At intervals of

14, 26, 34, 45, and 62 wk after t h e start of BBN exposure, both exposed a n d control animals were sacrificed, their bladders inflated with trypsin-EDTA solution (61, and their urothelial cells dissociated into a suspension of single cells, which were then deposited on Millipore filters for staining.

Instrumentation

Papanicolaou-stained bladder cells were scanned on the ACUity microscope system i n the Biomedical Sci- ences Division of Lawrence Livermore National Labo- ratory. The hardware a n d software components of

ACUity have been described elsewhere (211, so only a

brief description will be given here. A Leitz Ortholux microscope, configured for absorption microscopy, was

used with a 1OOx oil-immersion objective (NA = 1.32)

'Work performed under the auspices of the U.S. Department of Energy by the Lawrence Livermore National Laboratory and the Uni versity of California, San Francisco, under contract W-7405-ENG-48, and a grant from the National Bladder Cancer Project, grant CA- 23790.

(2)

RAT BLADDER IMAGE CYTOMETRY 455

and an 8 x ocular to provide images with spatial resolu-

tion of approximately 0.25 pm. The images were scanned with a Sierra Scientific vidicon camera a t a sampling density of 8 points per linear micrometer with a signal-

to-noise ratio in excess of 50 db. The video image was

digitized using a Quantex DS/20, were 16 consecutive frames were summed to further reduce noise and a 128- by-128 image was then transmitted over a high-speed digital link to a PDP 11/34A computer for analysis and storage.

Image Analysis

Three sets of images were scanned and analyzed a t each of the five time points (14, 26, 34, 45, and 62 wk after the start of carcinogen exposure). Each set contained about 50 well-preserved small or intermediate urothelial cells. The three groups were (1) randomly

selected cells from two or three control rats, (2) cells selected as morphologically abnormal from two or three test rats, and (3) randomly selected cells from two or three test rats. Group 2 was selected by a cytologist, while groups 1 and 3 were selected by someone not trained in diagnostic cytology. The resulting cell images were analyzed for morphologic features and the ex-

tracted parameters analyzed by the UCLA BMDP statis-

tics package (5).

When the picture-analysis program processed a picture, that file was read into computer memory and also copied to a reserve section of memory that is part of a memory-mapped display system. This system displays a picture with 256 gray levels and is partitioned by software to display four 128-by-128 pictures. Thus at any stage the operator could see up to four cells displayed.

Typical pictures are shown in Figure 1.

FIG. 1. Video images of cells used for image processing. Two panels

of 128 x 128 pixel images are displayed. The left (A) is a normal cell, while the right (B) is a n abnormal cell. Both are from the 45-wk time point. Images such as this were displayed for the operator to adjust the threshold in determining the nuclear and cytoplasmic boundaries. In A the cytoplasmic contour has been drawn in black, while in B the nuclear contour is displayed.

FIG. 2. Panel A (left) shows the corresponding three level transformation of nuclear brightness for the control cell shown in Figure 1A. Panel B (right) shows the three-level transformation on an abnormal cell. The examples shown illustrate the increased chromatin heterogeneity of the abnormal cells.

(3)

The analysis program has two stages: scene segmentation and feature extraction. In the scene segmentation stage the individual cell is located and the boundaries of its nuclear and cytoplasmic regions determined (Fig.

1). In the feature extraction stage, morphological mea-

surements (size, shape, texture, and stain) are made on the nuclear and cytoplasmic regions.

Scene segmentation. Although a variety of techniques are available for identifying the object boundaries in pictures (10,20), we chose human-assisted thresholding. This technique was selected because we found that no single algorithm was completely effective for the two boundaries per cell in all cells involved in this study. Through interaction with our high-quality display, it was quite simple, however, for a n operator to choose a threshold for scene segmentation.

Feature extraction. After the nuclear and cytoplasmic boundaries were determined, the following features were determined for both the nucleus and the cytoplasm: Size:

Shape: normalized perimeter2/area, bending en-

area, perimeter, nuclear aredtotal cell area ergy, mean absolute curvature, sphericity, ellipitcal eccentricity

total stain (i.e., total optical density), total staidarea

Stain:

On the nucleus we determined three measures of chromatin distribution:

heterogeneity, condensation, and margination

For those features whose definitions are not obvious, a

brief description follows.

Normalized perimeter2/area. This shape factor, some- times referred to a s the thinness ratio is defined as

P2A = (perimeter)2

471. x area

and takes on its minimum value.of 1.0 only for a circle (13).

Bending energy. This parameter represents a model for the amount of energy stored in a shape (2,221 and also takes a minimum value for a circle. Functionally, it is given by

where K(p) is the curvature of the shape contour a t point p and the computation is done for every point on the contour.

Mean absolute curvature. This shape parameter (2) is similar to bending energy and is given by

Sphericity. This parameter measures the ratio of the radius of a circle with area equal to the given contour to the radius of a circumscribing circle (18).

Elliptical eccentricity. This shape measure is based upon the idea of the eccentricity associated with a n elliptical contour, specifically the ratio of the major to the minor axis. We fit a n ellipse to the contour of the cell by assuming that the cell is a homogeneous mass and calculating the second moments of inertia about the center-of-mass. With these moments represented as a matrix

Mxx Mx,

the eccentricity is defined as the ratio of the larger eigenvalue of the matrix to the smaller eigenvalue.

Chromatin Measures

One of the most difficult problems in the quantitative description of cell images is finding satisfactory measures for nuclear chromatin distribution. Using conventional gray-level images of cell nuclei, we have been guided by the human linguistic descriptions of chromatin in quantifying texture. Two key linguistic concepts are chromatin condensation into clumps and the size of the clumps. Our method follows the subjective impres- sion that only three density levels, corresponding to white, gray, and black, are necessary to the perception of chromatin condensation. In a n absorbance image, black corresponds to highly compacted chromatin, white to cleared chromatin, and gray to diffusely distributed chromatin.

The transformation of the nuclear image from one with 256 gray levels to one with these three levels is accomplished by automated thresholding based upon percentages of the average nuclear brightness. As is shown in Figure 2, areas above the empirically deter-

mined 1.2 x [ mean brightness] are assigned black, areas

below 0.8 x [mean brightness] are assigned white, and

intermediate areas are assigned gray. This simple three- level partitioning of the nucleus was used to determine the degree of heterogeneity in the chromatin and the condensation of chromatin. The border of the nucleus was always identified in white, as in Figure lB, and was not used in the calculation of heterogeneity and condensation. A detailed description of these features is in preparation (23), and only a brief summary follows.

Heterogeneity. To assess the chromatin distribution, the total number of black or white labeled points (see Fig. 2) is divided by the total nuclear area. If the nucleus is homogeneous (i.e., gray), then this parameter is zero. If the nucleus is completely heterogeneous (every point being either black or white) this parameter is 1.0.

Condensation. To assess the size of condensed chro-

(4)

457 in the nucleus through a square mesh eight pixels (ap-

proximately one micrometer) on a side. The absolute

difference between the number of black points and the number of white points within a mesh window is determined and the sum of such absolute differences com- puted. This sum is then divided by the area for normalization. If the nuclear chromatin is finely granular (like salt and pepper), then the value of this parameter approaches zero. If the nuclear chromatin is made up of coarse grains (like marbles), then the value of this parameter approaches 1.0.

Margination. By this term we refer to the increase of

chromatin at the nuclear boundary. To quantify this

concept we measure the stain content a t every point in the nucleus using the original 256 gray-level image and then measure a quasi-radial moment of inertia about the center-of-mass (23). Using a circular nucleus of unit radius a s a n example, we find that if all the mass is concentrated at the center we get a value of zero for the parameter. For a uniform, homogeneous distribution of stain mass we get a value of 0.33. For all of the mass

concentrated around the rim of the circle, the parameter

takes on a value of 1.0.

Conventional Classification

To test the cytologist’s observation that the early and

late abnormal cells seem to represent two distinct populations, the former with features of repair and the latter with features of neoplasia, the sets of selected abnormal cells were displayed in random blind order to the cytotechnologist and ranked according to the perceived time order of the set. In this way we were able to determine if human perception could identify differences between the early and late abnormal cells based only on the images actually stored for computer analysis.

In addition, 26 images from each group of randomly selected test cells (i.e., BBN-exposed cells) were dis- played for the cytologist to evaluate. Each cell was then scored according to conventional cytopathology norms as normal, abnormal, or borderline.

RESULTS

The cytologist reviewed all of the stored digital images

and concluded that the sample preparation of the control

cells at the 62-wk time point had differed from that a t other time points, in that the cells appeared pale and

possibly overtrypsinized. As such the control cells from

the last time point (wk 62) were eliminated from all statistical analysis.

In the blind ranking of the set of abnormal cells based on the digital images, the cytotechnologist ordered the files 34, 26, 14, 45, and 62 wk. This incorrect transposi- tion of the 14th and 34th wk time points suggests that, while there were clear differences distinguishing the malignant cells a t the last time points (45 and 62 wk), the differences between the reactive hyperplastic cells at 14,26, and 34 wk were less obvious.

Contour fitting with operator-assisted thresholding worked well for determining the nuclear boundary. It

worked less well for finding the cytoplasmic boundary owing to the variability and low contrast in cytoplasmic stain density, and vacuoles in the cytoplasm. For example, Figures 1A and 2A are of the same cell. The computer-determined cytoplasmic contour showed a slight indentation a t 10 o’clock in Figure lA, which is not correct, as seen in Figure 2A. As such, the statistics based on cytoplasmic parameters had large standard deviations and variability between groups that seemed unrelated to time or the neoplastic process. Because of these errors associated with the cytoplasmic statistics, we did not analyze them further.

As a n example of the analysis procedure we show in Ta-

ble 1 the numerical values of some of the parameters for

each of 24 cells. As each set of features was extracted from a given cell, the vector of features was appended to a feature file. After the cells of a sample have been analyzed, the resulting feature file contained the feature vectors for each cell, ready for data reduction. A sample statistical summary based on means, variances, and covariances for

a population of 66 cells is given in Table 2.

The two statistical tests in the UCLA BMDP package that were used extensively in this project were P3D, the

multivariate t-test, and P3S, the nonparametric Mann-

Whitney rank order analysis. Using these two we searched for that subset of features that best described the differences among the random, abnormal, and control groups a t each time point.

Nuclear Stain and Size Parameters

Figure 3 shows the change in nuclear area with time for the control cells and the selected abnormal cells.

The nucleus was larger in abnormal cells at the 14-, 34-, and 45 wk time points. The largest nuclei were found a t 62 wk in the abnormal cells. The small varia-

1 4 _I ₁ _I

1

1 3

1 t

1 091 1 1 I I I I

I

10 20 30 40 50 60 70 Time (wk)

FIG. 3. Plot of nuclear area a s a function of time for control (I ) and

abnormal cells (A). Abnormal cells were larger t h a n control cells a t all time points, but not significantly a t the 26-wk point. P values for differences between paired groups using the Mann-Whitney ranks test were 0.05,0.0146, and 0.0037 a t 14, 34, and 45 wk, respectively. Means

(5)

Table 1

Table of Nucleus Feature Values for the First 24 out of 66 Cells i n a Feature Data File Cell No. Area Perimeter P**2/a

1 969.00 2 945.00 3 524.50 4 1,126.50 5 1,046.50 6 839.50 7 763.00 8 699.00 9 748.00 10 843.50 11 684.00 12 1,011.00 13 602.50 14 1,078.00 15 1,081.50 16 688.00 17 1,433.00 18 2,189.00 19 663.50 20 1,063.00 21 785.00 22 829.00 23 847.50 118.23 1.15 119.40 1.20 88.18 1.18 127.64 1.15 122.33 1.14 115.98 1.28 107.74 1.21 100.43 1.15 103.25 1.13 112.81 1.20 98.91 1.14 121.40 1.16 109.15 1.18 127.05 1.19 129.64 1.24 102.91 1.22 143.88 1.15 184.85 1.24 97.50 1.14 128.23 1.23 106.08 1.14 108.57 1.13 112.33 1.18

Bend nmac Spher ecc N/Cratio Stain 1.62 2.02 1.68 1.62 1.81 2.07 1.99 1.66 1.74 1.90 1.58 1.74 1.74 1.97 1.95 2.17 1.88 2.13 1.53 2.22 1.57 1.56 1.91 1.18 1.26 1.15 1.17 1.15 1.18 1.22 1.15 1.15 1.20 1.18 1.17 1.17 1.18 1.20 1.25 1.18 1.19 1.17 1.26 1.15 1.18 1.16 0.28 1.17 0.33 1.50 0.31 1.38 0.80 1.39 0.26 1.20 0.24 1.96 0.46 1.47 0.44 1.13 0.20 1.57 0.52 1.47 0.70 1.62 0.27 1.50 0.65 1.49 0.22 2.14 0.04 1.93 0.55 1.32 0.76 1.35 0.58 1.93 0.58 1.19 0.54 1.89 0.07 1.38 0.45 1.07 0.18 1.49 0.50 0.32 0.21 0.46 0.36 0.40 0.35 0.52 0.21 0.32 0.30 0.47 0.40 0.63 0.34 0.22 0.52 0.35 0.28 0.36 0.24 0.48 0.44 2,111.57 2,397.61 1,780.66 2,054.04 2,662.89 1,867.96 1,261.61 1,283.34 1,546.63 2,223.73 1,252.03 2,032.07 3,054.91 1,558.07 1,986.18 987.23 2,558.64 3,845.4 1 1,447.74 1,706.36 1,224.78 2,230.49 1,816.15 1.321.46 S t a i d p n t Hetero Clump 1.25 0.71 1.45 0.74 1.67 0.66 1.09 0.69 1.51 0.67 1.22 0.57 0.89 0.67 1.27 0.73 1.13 0.68 1.45 0.62 1.26 0.62 1.47 0.66 1.16 0.65 0.85 0.62 1.07 0.57 0.98 0.73 0.80 0.67 1.41 0.65 1.50 0.59 1.17 0.65 1.10 0.71 1.50 0.74 0.97 0.62 1.18 0.69 0.74 0.64 0.49 0.79 0.72 0.76 0.77 0.81 0.51 0.66 0.59 0.75 0.51 0.70 0.76 0.61 0.73 0.71 0.58 0.77 0.65 0.75 0.56 0.59 24 796.00 105.74 1.12 1.62 1.17 0.17 1.10 0.32 Table 2

Descriptive Statistics From 66 Cells Chosen at Random From the Rats at A g e 42-45 wk”

401

411

421 f131 f141 f151 461 471 f181

491 4101

f l l l i Means 947.92 Std deviation 252.00 Std dev of means 31.02 CV 26.58% Correlations f1Ol 1.000 411 0.990 f121 0.147 f131 0.303 f141 0.084 f151 0.079 f16l 0.171 -0.002 0.690 -0.055 4101 0.024 4111 0.400 ffOl f171 f191 481 116.50 1.16 1.72 1.18 0.40 1.37 0.39 1953.16 1.28 0.68 0.67 15.18 0.04 0.17 0.03 0.20 0.24 0.11 553.07 0.24 0.05 0.10 1.87 0.00 0.02 0.00 0.02 0.03 0.01 68.08 0.03 0.01 0.01 13.03% 3.13% 9.94% 2.53% 49.35% 17.49% 28.29% 28.32% 18.41%7.09% 14.62% 1.000 0.233 0.379 0.135 0.048 0.238 -0.002 0.685 -0.079 -0.003 1.000 0.868 0.573 0.762 0.126 0.007 -0.037 -0.079 -0.301 1.000 0.643 1.000 -0.028 0.161 1,000 0.673 0.281 -0.120 1.000 -0.090 -0.218 -0.033 0.087 -0.158 0.107 0.186 -0.149 - -0.229 -0.010 -0.025 -0.490 0.160 0.034 0.189 0.040 1.000 0.045 1.000 -0.119 0.302 1.000 0.104 0.121 0.231 1.000 0.444 0.152 0.236 0.171 -0.125 0.049 -0.047 0.134 -0.227 0.066 1.000 .

411

f121 431 441

451

mi 471

481

f191 mi mi

VOl,

nucleus area; 411, nucleus perimeter; 421, nucleus p**2/area; f13], nucleus bending energy; 441, nucleus mean abs. curvature; f151, nucleus sphericity; Q6], nucleus eccentricity; f171, N/C ratio; 481, total nucleus stain; f191, nucleus s t a d p o i n t ;

f1

101, nucleus chromatin hetero; 4111, nucleus chromatin clumping.

tions noted in the control groups for the first four time points suggests that the control nuclei at 62 wk also would have been significantly smaller than the abnormal nuclei. Using the nonparametric Mann-Whitney

ranks test, the P values for differences between the

paired samples a t 14, 34, and 45 wk were 0.05, 0.0146, and 0.0037, respectively.There was no difference in nu- clear area at 26 wk. A similar temporal pattern of differences was noted in the nuclear perimeter measure.

Total nuclear stain content, shown in Figure 4, had a slightly different pattern. Differences between abnor-

mal and control samples were not significant at 26 and

34 wk, but a P value of 0.028 was observed at 14 wk and

of less than 0.0001 was observed a t 45 wk.

Nuclearhotal-cell area ratio (N/C) was of low “resolv-

ing” power, probably because of the uncertainties in the cytoplasmic measurements. The same general pattern for N/C ratio as in the nuclear area measurements was

seen, however, and a t 45 wk the P value was 0.0006.

Taken together these measurements show that both the reactive hyperplastic abnormal cells present at 14 wk and the malignant abnormal cells present at 45 and

(6)

459 1.5 >

-

1.4 w 1.3 2 3 _I _I _I _I _I _I

_I

2 6 I I I I I I - 7 - - Time I w k ) Time (wkl

FIG. 4. Total nuclear stain a s a function of time for control (0) and

abnormal cells (A). Abnormal cells had darker nuclei a t all time points than control cells, although t h e difference was not significant a t 26 and 34 wk. P values for the Mann-Whitney ranks test were 0.028, 0.29, 0.49, and 0,0001 for t h e four time points consecutively. Means i 1 standard deviation a r e shown.

FIG. 6. Plot of bending energy a s a function of time. Abnormal cells

( A ) had more indentations in their nuclear contours than did control cells (0) a t 14, 36, and 62 wk. P values for the Mann-Whitney ranks test a t 14 and 36 wks were 0.0008, and 0.0004, respectively. Means -t

1 standard deviation are shown.

ation, and two measures of this-bending energy and

perimeter squared/area-showed similar patterns. The

I

T

I 1

bending energy data are presented in Figure 6. The control contour of the abnormal cells was generally more irregular than the controls, although this was not significantly so a t the 26- and 45- wk time points. Large bending energy was observed in 14-, 34-, and 62-wk abnormal cells, and there was large variation in the cells. P values for the rank test were 0.0008, and 0.0004

at 14 and 34 wk, respectively. The perimeter squared /

area was larger in abnormal cells a t 14 and 45 wk and

I

,

I

,

I

1

was also high a t 62 wk, but less dramatically than the

1.2 bending energy statistic.

0 10 20 30 40 50 60 70

Time Iwkl Chromatin Distribution Parameters

FIG. 5. Plot of nuclear elliptical eccentricity a s a function of time.

Abnormal cells ( A ) had more elongated nuclei a t all time points than control (0). Most elongation was observed in t h e oldest test rats. Means

+

1 standard deviation are shown.

All three texture parameters differed significantly be-

tween test and control cells at all tirne points. The data

on nuclear heterogeneity are illustrated in Figure 7. Abnormal cells had chromatin that was more heterogeneous. A similar pattern was observed for chromatin condensation, with abnormal cells having larger clumps 62 wk had enlarged, darker nuclei. This finding is con- of dark chromatin than control cells, as is shown in

sistent with the pathologist’s verbal description of nu- Figure 8. These observations are consistent with the

clear enlargement and hyperchromasia. cytopathologist’s verbal description of increased chro-

Nuclear Shape Parameters

The elliptical eccentricity of the nucleus was larger in abnormal cells a t all time points except the 45-wk point, and the results are shown in Figure 5 . The differences at the first three time points resulted in ranks test P

values of 0.04, 0.378, and 0.001. The high degree of ellipticity in the abnormal cells a t 62 wk is not surpris- ing, since some areas seen in histologic sections a t this time point had very dysplastic cells with marked polar orientation (6).

matin condensation in abnormal cells. Interestingly, these measures (with the possible exception of condensation) did not show a difference between the chromatin in the early 14-wk (hyperplastic) cells and the late ( 2 45 wk, malignant) abnormal cells.

The nuclear margination measurement of chromatin

distribution also showed significant differences between abnormal and control cells at all time points, and the results are shown in Figure 9. At all time points the margination was greater in control cells than in abnormal cells. The abnormal cells were relativelv constant

Irregular indentations of the nuclear perimeters in over time in the degree of margination, while the control

(7)

460 0 70 I I I I I I

t

1 t

1

0 5 1 I I I I I I 0 10 20 30 40 50 60 70 Time (wk)

FIG. 7. Plot of heterogeneity in chromatin as a function of time. The larger the value of this statistic, the wider the distribution of densities in the picture points of the nuclei. Abnormal cells (A) had greater heterogeneity in the chromatin density a t all time points than did control cells (0). P values for the Mann-Whitney ranks test were 0.0001, 0.05, 0,0001, and 0.0001 a t the four time points consecutively. Means f 1 standard deviation are shown.

0.5 I _I _I _I _I _I

T I

8

0.34L”J

0.3 0 10 20 30 40 50 60 70

Time (wk)

FIG. 8. Plot of condensation of chromatin in cells as a function of time. The larger the value of this statistic, the more coarse the grains of chromatin. Abnormal cells ( A ) a t all time points had more coarsely granular chromatin than did control cells (0). P values for the Mann- Whitney ranks test a t all four time points were, consecutively, 0.0001,

0.001, 0.0001, and 0.05. Means i 1 standard deviation are shown.

The finding of greater “margination” in the control cells is in conflict with the cytopathologist’s description of greater margination in abnormal cells. It suggests that, while this statistic is well suited for differentiating abnormal and control cells, it does not reflect what the pathologist calls “margination.” In view of the fact that

a value of 0.33 is expected for this parameter when the

chromatin is uniformly distributed, values less than 0.33

in the abnormal cells suggest that this parameter is being dominated by a n increased dark central mass in the normal nuclei, such as the nucleolus. The idea that the nucleolus is increased in abnormal cells is consistent with cytopathological description. Examining the cells

0.4 _I _I _I 6 0.32 E

K

0 2 8

K

0’32L

0 2 8

W

l

T 0.2

I

I I I I I I 10 20 30 40 50 60 70 Time ( w k )

FIG. 9. Plot of nuclear margination as a function of time. A value of 0.33 in this parameter corresponds to a completely even distribution of stain intensity. With this means of calculating margination, abnormal cells ( A ) had less margination than control cells (0). P values were 0.0001. 0.0001, 0.014, and 0.016 a t the four consecutive time points using the Mann-Whitney ranks test. See text discussion for how the presence of nucleoli may influence calculation of this measure. Means k 1 standard deviation are shown.

shown in Figure 2 illustrates the problems encountered

in quantitating margination. The normal cell (2A) has a

large nucleolus on the nuclear border, and the abnormal

cell (2B) has marginal nucleoli, a central nucleolus, and

some margination a t 8 o’clock. These images are too

complex for the simple definition of margination that we tried to apply. This experience with the “margination” parameter indicates some of the pitfalls in trying to match a computer formula to a verbal description of

nuclear morphology. Only with several iterations of

matching quantitative results with verbal descriptions is it possible to develop a n appreciation for the discrimi- natory power of a particular test.

Randomly Selected Cells From Treated Animals At each time point 50 cells were selected a t random from the test animals. These stored cell images were reviewed by the cytotechnologist, who classified the cells as normal, abnormal, or borderline. The fraction of cells classified a s normal was 0.27, 0.54, 0.16, and 0.11, and

0.11 a t 14, 26, 34, 45, and 62 wk, respectively. The

occurrence of the highest frequency of normal cells at 26

wk reflects the fact that at wk 14 there were many reactive hyperplastic cells and that neoplastic cells did not occur in great number until the later time points.

As might be expected, the parameters measured by

image analysis for the set of randomly chosen, exposed cells were between the value obtained for the control set and the abnormal set. Randomly chosen cells were most like control cells a t 14 and 26 wk and most like abnormal cells a t 45 and 62 wk, in correspondence with the

cytopathologist’s scoring of the relative frequency of nor-

mal cells in this group. A typical parameter distribution

(8)

z 6 2 - c m

-f

f 5.8 5.4 7 0 I _I _I _I _I ₁ - -

-

I 1 1 1 1 I OO 10 20 30 40 50 60 70 Time ( w k )

FIG. 10. Plot of chromatin heterogeneity for the randomly chosen

cells from treated animals (0) compared with cells from control animals (0) and cells selected as abnormal (A). The value of this statistic increased from near control values a t the early time point to near abnormal-cell values a t the last time point. This increased reflects the increase in atypical cells as a fraction of the total seen in the popula- tion of randomly selected cells. Means i~ 1 standard deviation are shown.

data on control and abnormal cells shown in this figure are the same a s that in Figure 7 and are presented here as a guide in judging the random cell data.

DISCUSSION

Dissociated rat bladder cells provide excellent cytologic material with which cell morphology algorithms can be developed, tested, and used to characterize populations. By using a n animal model, we were able to sample repeatedly from a controlled neoplastic process and develop a detailed data set on the time course of cytologic changes. We were able to make statistically precise measurements of nuclear features, but not of cytoplasmic features. The cells derived from rats showed the same pattern of morphologic changes associated with human bladder cancer and argue well for the use of this animal model in studying carcinogenesis.

At wk 14 the cells presented features of regenerative hyperplasia. This included some increase in nuclear size, stain, and indentations, and a marked increase in chromatin heterogeneity and condensation.

At wk 26 the cells were nearly normal. While the cytologist selected abnormal cells, these were not meas- urably different from untreated cells for parameters related to nuclear size and shape. These cells retained their abnormal chromatin pattern, however. In randomly selected cells from test animals, the largest fraction of normal cells was found at this time point.

The wk 34 abnormal cells showed differences from control that were similar in magnitude to those seen a t wk 14. In fact, a paired comparison between the features of selected abnormal cells a t 14 and 34 wk shows no features that differentiate these two populations. This

461 observation implies that the features used cannot discriminate between hyperplastic cells (the 14-wk abnormals) and the earliest neoplastic cells (the 34-wk abnormals). Interestingly, in the blind comparison, the cytologist also confused the stored abnormal cell images from these two time points.

At wk 42 the cytologist noted cells with features nor- mally associated with malignancy, and in the randomly

selected cells from test animals only 11% were scored as

normal. The abnormal cells a t this time point were characterized by a n increase in nuclear area, a marked increase in nuclear stain, no difference from control in eccentricity of bending energy, and highly abnormal chromatin staining.

The abnormal cells a t wk 62 were judged most bizarre by the cytologist and showed highly atypical features for all the statistics presented, including eccentricity and bending energy. While there was no specifically matched control population for these cells, the statistics on the control cells a t the other time points were suffi- ciently constant to suggest that all of the differences seen a t wk 62 reflect the neoplastic process rather than aging.

Several attempts by others have been made to quantitate the cytologic changes in urothelial cells as a means of improving diagnostic cytology. These include the measurement of DNA content by flow cytometry (7,11,12, 14,16), and the analysis of images of Papanicolaou- stained cells (8,9,15). Both approaches have used human cytologic material and have shown significant differences between normal cells and frankly malignant cells. Both approaches have been less than ideal, however, in finding reliable means of identifying cells from low- grade lesions andor being able to relate their measures to conventional cytopathologic description.

These studies merit particular discussion. First is the

series of papers by Koss and co-workers using TICAS, a

scanning system similar to ours (8, 9, 15). Our study differs principally from theirs in our use of texture measures that we believe relate to the cytopathologist's judg- ment of chromatin compaction and our use of these quantitative tools on rat cells in graded model tumor system.

In their studies, as in ours, chromatin texture features had strong discriminating power between the sets of control and abnormal cells. They were able to define a combination of measurements which, if used together, could classify cells as normal or malignant. Using this classification scheme, two classes of cells were identified, typical I and typical 11, which fell in between nor- mals and abnormals. Because the parameters used were only slightly related to convention cytopathologic descriptions, it has been hard to interpret these grades of

atypia in terms of either standard cytology or in terms

of the neoplastic process. We intend our approach to link more closely the measured cell statistics to the pathology diagnosis and to the temporal pattern of neoplastic progression.

Flow cytometry studies by Scandinavian groups (3,4, 14,16) and by Melamed's group (7,121 in New York have

(9)

462

linked quantitative

DNA

measurements to cytologic

grading of human bladder tumors. In both studies, tumors of increasing grade showed increasingly abnormal

histograms. Grade I and

I1

tumors, however, were only

minimally different from controls. These studies suggest that our image analysis approach would be improved if

we used a quantitative

DNA

stain, such as the Feulgen

reaction, and were able to perform measurements of total nuclear stain intensity more precisely while retain- ing the important parameters associated with chromatin texture.

Finally, it is worth noting that there have been many other changes in bladder tumor cells described besides

those of nuclear morphologic and

DNA.

These include

lost of

ABO

antigens (191, and the expression of in-

creased

NADH

diaphorase (17). The latter, in particular,

seems to discriminate between neoplastic and hyperplastic responses at a very early stage in the rat bladder cancer model. Enzyme histochemical stains or immuno- cytochemical staining using the precision of monoclonal antibodies may offer staining methods for the cytoplasm that would complement the studies reported here on nuclear morphology.

ACKNOWLEDGMENTS

We would like to acknowledge the advice and help of our colleagues Ms. Debby Bennett, Mr. Tom Slezak, and Dr. Brian Masall. 1. 2. 3 . 4. 5. 6. 7 LITERATURE CITED

Beyer-Boon ME, De Voogt HJ, Van der Veble EA, Brusse JAM, Schalsery A: The efficacy of urinary cytology in the detection of urothelial tumors. Urol Res 6:3-12, 1978.

Bowie J, Young I: An analysis technique for biological shape- 11. Acta Cytol, 21:455-464, 1977.

Farsund, T Preparation of bladder mucosa cells for micro-flow

fluorometry. Virchows Arch [Cell Pathol.] 16:35-42, 1974. Farsund T: Cell kinetics of mouse urinary bladder epithelium. Virchows Arch [Cell Pathol.] 21:279-298, 1976.

Hill M: BMDP User’s Digest: A Condensed Guide to the BMDP

Computer Programs. BMDP Statistical Software. Dept. of Mathe- matics, UCLA, 1979.

King EB, Vanderlaan M, Jensen RH, Kromhout L, Hoffman J:

Morphologic changes in rat urothelial cells during carcinogenesis: I. Histologic and cytologic changes. Cytometry 5:447-453, 1984. (this issue)

Klein FA, Melamed MR, Whitmore WF, Herr HW, Sogani OC,

Darzynkiewicz Z: Characterization of bladder papilloma by two- parameter DNA-RNA flow cytometry. Cancer Res 42:1094-97, 1982.

8. Koss LG, Bartels PH, Bibbo M, Freed SZ, Taylor J, Weid GL:

Computer discrimination between benign and malignant urothelial cells. Acta Cytol 19:378-391, 1975.

9. Koss LG, Bartels PH, Sychra JJ, Wied GL: Diagnostic cytologic sample profiles in patients with bladder cancer using TICAS- system. Acta Cytol 22:392-397, 1978.

10. Lester J, Williams H, Weintraub B, Brenner J : Two graph search-

ing techniques for boundary finding in white blood cell images. Comput Biol Med 8:293-308, 1978.

11. Levi PE, Cooper EH, Anderson CK, Williams RE: Analysis of DNA content, nuclear size, and cell proliferation of transitional cell carcinoma in man. Cancer 23,51074-1085, 1969.

12. Melamed MR, Darzynkiewicz Z, Traganos F, Sharpless T Nucleic

acid content and nuclear chromatin structure of bladder cell cul- ture lines as studied by flow microfluorometry, Cancer Res 37:1227-1231,1977.

13. Navarro EF: The P2A function, circularity, and elipticity of digital

images and polygonal curves. M.S. Thesis, Cornell University, August, 1979.

14. Petersen T, Larsen JK: Flow cytometry in bladder washings, Ugeskr Laeg 140:161-164, 1978.

15. Sherman A, Koss LG, Adams S, Schreiber K, Moussouris HF,

Freed SZ, Bartels PH, Wied GL: Bladder cancer diagnosis by

image analysis of cells in voided urine using a small computer. Anal Quant Cytol 1981, pp 239-249.

16. Tribukait B, Gustafson H, Esposti P: Ploidy and proliferation in human bladder tumors as measured by flow cytofluorometric DNA analysis and its relations to histopathology and cytology. Cancer 43:1742-1751, 1979.

17. Vanderlaan M, Fong S, King EB: Histochemistry of NADH diaphorase and y-glutayltranspeptidase in rat bladder tumors. Car- cinogenesis 3:397-402, 1982.

18. Wadell H: Sphericity and roundness of rock particles. J Geol

41:310-331, 1933.

19. Weinstein RS, Alroy J, Farrow GM, Miller AW, Davidsohn I: Blood group isoantigen delection in carcinoma in situ of the urinary bladder. Cancer 43:661-668, 1979.

20. Yasnoff W, Mui J, Bacus J Error measures for scene segmentation.

Pattern Recognition 9:217-231, 1977.

21. Navarro EF: The P2A function, circularity, and elepticity of digital images and polygonal curves. M.S. thesis, Cornell University, Au- gust, 1979.

22. Young IT, Gledhill BL, Lake S, Wyrobek A J Quantitative analysis

of radiation-induced changes in sperm morphology. Anal Quant Cytol 4(3) :207-216, 1982.

23. Young I, Walker, J Bowie J : An analysis technique for biological shape I. Information Control 25:350-370 1974.

24. Young IT, Verbeek PW, Mayall BH: the analysis of biological texture, I: The distribution of nuclear chromatin. Comput Graph- ics Image Process (in preparation).