computerized nuclear morphometry as an objective method for characterizing human cancer cell...

[CANCER RESEARCH 38, 4688-4697, December 1978]0008-5472/78/0038-0000$02.00

Computerized Nuclear Morphometry as an Objective Method forCharacterizing Human Cancer Cell Populations1BjOrn Stenkvist,2 Sighild Westman-Naeser, Jan Holmquist, Bo Nordin, Ewert Bengtsson, Jan Vegelius,Olle Eriksson,andCecil H. Fox3Department of Clinical Cytology, University Hospital, S-750 14 Uppsala, Sweden (B. S., S. W-N.J; Departments of Computer Science (J. H., B. N., 0. E.j,Physics (E. B.), and Statistics (J. V.1, Uppsala University, Uppsala, Sweden; Department of Pathology, Karolinska Institute, 104 01 Stockholm, Sweden (C.H. F.); and Laboratory of Biochemistry, Division of Cancer Biology and Diagnosis, National Cancer Institute, Bethesda, Maryland 2@X@14(C. H. F.)

ABSTRACTA new method for measuring differences in nuclear

detail in chromealum galbocyanin-stainednucleiof cellsfrom humanbreast cancerswas comparedwith conventional subjectivegradingand classificationsystems.Thenew method,termed computerizednuclearmorphometry(CNM), gives a multivariatenumericalscore that correbateswellwithnuclearatypiaandgivesa higherreproducibility of classification than do subjective observationswith conventional histologicalpreparations. When 100individualnuclei from each of 137 breast cancers wereexaminedbyCNM,therewasa broadCNMscorevariationbetween patients but a good reproducibilityfor eachtumor. When different parts of the same tumor weresampled, there was good reproducibilitybetween sampIes, indicatingthat some breast cancers at least aregeometricallymonocbonal.When these cancers werecompared by the grading systems of WHO and Black,correlationsof 0.43 and 0.48, respectively,were found.There was a poor correlationbetweenCNM and classifications of tumor type, but in general there were highvalues for CNM in medullarytumors and low values inmucoustumors. Correlationsbetween CNM and tumorprogressionand prognosisawait futurestudyof patientsparticipating in the study.

INTRODUCTION

Human breast cancers are diagnosed on the basis of theappearance of tissue or cells removed from the patient andprepared for microscopic evaluation. In addition to thediagnosis of cancerous tissue, most pathologists also attempt to classify the histological origin of the tumor and toestimate the gradeof the tumor. While the significanceof subjective grade estimates for prognosis has been debated (8, 22, 24), there have been a number of such systemsproposed, i.e. , those of Bloom and Black (4, 5). A numberof classification systems have also been suggested, i.e.,those of WHO, Ackerman, and Armed Forces Institute ofPathology(1,21,25).

In an effort to improve reproducibility of grading of breastcancers, we have used digitized image analysis of the nucleiof breast cancer cells that were stained with chrome alum

I This work was supported by NIH-National Cancer Institute Contract NOlCB-53968 through the Breast Cancer Task Force.

a To whom requests for reprints should be addressed.3 On leave from Laboratory of Biochemistry, Division of Cancer Biology

and Diagnosis, National Cancer Institute, Bethesda, Md.Received December 27, 1977; accepted September 15, 1978.

galbocyanin. The results of this objective grading systemare compared with the results of subjective grading andclassification. In addition, we have succeeded in obtaininga quantitative expression of nuclear pleomorphism thatallows the determination of the reliability of our samplingmethods and also allows the examination of the biologicalhomogeneity of a breast cancer.

MATERIALSAND METHODSPatients

Cellular material was obtained from 162 of 181 consecutive breast cancers removed from patients in 4 Swedishcounties during a period of 5 months. The patients wereunusually well characterized as a part of a larger projectcorrelating epidemiology and morphology of breast cancer.A detailed history, physical examination, and questionnairefor epidemiobogical information were administered to eachpatient, and blood samples were drawn both before andafter surgery for use in hormone and other assays. To date,the patients' course has been followed for 2 years, withcollection of clinical course observations continuing.

Specimen Preparation

Each tissue specimen was taken directly to the laboratoryfollowing its removal. The fresh tumor was biopsied first byfine-needle aspiration with a 22-gauge needle and a 20-mIsyringe in a syringe holder. This method produces singledispersed cells that are well suited to cytological study.Following aspiration biopsy the tumor was divided, the sizeof the tumor was measured, and portions were taken fordetermination of estrogen receptors and histopathology.Imprint specimens were obtained from the cut surface ofthe tumor by gently pressing a microscope slide against thesurface. All cytological specimens were fixed immediatelyin fresh Carnoy's solution and stained with chrome alumgallocyanin (9, 10). Of the 179 tissue specimens, adequateand complete cytological material was obtained from 162.Of these 137 specimens have been analyzed by CNM.4

For determination of the biological homogeneity of thetumors, fine-needle biopsies were taken from different partsof the same tumor in 14 cases. For demonstration ofcongruence between imprint and needle biopsy specimens,fine-needle biopsies and imprint specimens were obtainedfrom different parts of the same tumor and CNM valueswere compared.

4 The abbreviation used is: CNM, computerized nuclear morphometry.

CANCERRESEARCHVOL. 384688

on July 27, 2015. 1978 American Association for Cancer Research. cancerres.aacrjournals.org Downloaded from

Computerized Nuclear MorphometrySelectionof TumorCell Nuclei

Cytological preparations were carefully reviewed , and100 nuclei from each specimen were chosen for recording.Only typical cancerous epithelial cells with undamagednuclei were chosen. Each subsequent cell was selected forrecording by a standard process to avoid the operator'smaking subjective decisions of cell selection. At the sametime that each cell image was digitized, a color photograph(Ektachrome Film; Eastman Kodak, Stockholm, Sweden)was made in white light so that the selected field of visioncould be checked against the computer image.

Image Recordingand PreprocessingA diode array scanner coupled with a Leitz Orthoplan

microscope (E. Leitz, Wetzlar, Germany) was used to recordthe cell images. The details of the scanning system havebeen described elsewhere (2, 13) and will only be summarized here. The diode array scanning microscope wasdeveloped by Aslund and coworkers (17, 18) at the RoyalInstitute of Technology, Stockholm, Sweden and is not yetcommercially available. It consists of a linear array of 256diodes that measure the light in the image. The image isdivided into narrow bands by a swinging prism device, andeach of these image bands is measured by the diodes. Thesignal from each diode is digitized and transmitted to thecomputer interface and thence to a central processing unit.The image recording work was done with a PDP-8 minicomputer. The subsequent image processing was done with a

Chart 1. Hardware configuration. CPU, central processing unit; DEC, Digital Equipment Corporation.

PDP 11-55 (Digital Equipment Corporation, Maynard,Mass.). The light source for the microscope was fitted witha 570-nm interference filter (30-nm bandwidth) for scanning. Rectangular fields were scanned with a final spatialresolution of 0.2 @mwith a maximum image size of 256 x256 image points and 6 bits of gray-level resolution. Thehardware is diagrammed in Chart 1.

The digitization process for each image was followedwith a Tektronix 4010 (Tektronix Inc., Beavertown, Oreg.)display. The nuclear images were isolated from their surroundings by a rectangle defined by a cross-hair cursor andstored for further processing together with a gray levelcorresponding to the nuclear border. These gray levelswere determined from the gray-bevelhistogram of the visualdisplay by the use of an interactive graphic computertechnique previously developed (3, 20). All values above thenuclear border threshold were initially considered as belonging to the nucleus.

The images were then cleaned by local image processingoperations (2, 13). After this preprocessing image plotsshowing the border of each nucleus were visually reviewed,and images not clearly defined were excluded from thepopulation. This occurred in less than 0.2% of the cases,and in no instance were more than 10 nuclei from any onetumor excluded.

Mathematical Methods to Describe Nuclei and NucleiPopulations

The mathematical procedure for describing the numerical

RIN!COSPUTCSpop @s/r

DECEMBER 1978 4689


B. Stenkvist et a!.

features of a tumor cell population obtained from a breastcancer was structured into (a) parameters describing thecell nuclei and (b) parameters describing the entire cellpopulation based on the parameters extracted from thesingle nuclei.

Numerical Features of Nuclei

These parameters were subdivided into 3 categories,namely, into parameters describing density, shape, andtexture.

Density. The density parameters were derived from thenuclear gray-level histograms, in which h is defined as thenumber of image points in the region attaining gray-level iand i= 0 ...63 so that:

AREA=@h (A)

EXTINC=@ [Iog(64 log(i)}.h,

AVLEVEL =@ ih/AREA

VARLEVEL =@ (i@ AVLEVEL2h/AREA

ENTROPY = @h, .109(h)/AREA

The AREA parameter is the area over the region. EXTINC isthe amount of light absorbed by the region. AVLEVEL is theaverage gray-level over the region and VARLEVEL is thecorresponding variance. ENTROPY is a measure of theinformation content of the gray-level histogram.

Shape.TheshapeparameterswerebasedontheworkofGranlund (11) from which we developed computer subroutines for extracting shape descriptors from Fourier expansion of closed contours.

A point moving around a closed contour at constantspeed generates a complex function of time u(t. Thisfunction is periodic and can be expanded in a Fourierseries.

u(t) = ,,@ a,,e''

a, =@ f211u(t) .e'dt (G)2ir

assuming that the period of u is [email protected] coefficients a,, arecomplex and were calculated by the fast Fourier transformalgorithm after the closed contour had been interpolatedinto 512 sampled time points and moved to the origin of thecomplex plane followed by a normalization according tosize.

A truncated Fourier series

u(t) = ,, @m@ e@' (H)

was used to approximate the contour. The first 2 coeff I-cients determine an ellipse that approximated the originalcontour (in least square sense). The major axis MAJAXIS ofthis ellipse is Ia,I + Ia@,@,and the minor axis MINAXIS is lla@l1a,ll.

Descriptors based on the Fourier series must be independent of size, rotation, and orientation of the contour tobe useful as cell shape descriptors. Position and sizeindependence are obtained by the normalizations. Sincerotation through angle 0 affects the coefficients by a multiplicative factor e,the absolute values a,,f are independentof rotation of the contour. Equal contours with differentorientation (mirror images) give the Fourier coefficients a,,and a,,. Shape descriptors containing a,, and a@,,symmetrically will give orientation independence. From the Fouriercoefficients we calculated:

MAJTOMIN = a_,( + la,l (I)ffa_,fla,lI

BDYVAR (m, p) = (a,2 + &,2) (J)

MAJTOMIN is an elongation measure and BDYVAR is a(B) measure of the boundary variations in a part of the fre

quency spectrum.(C) Texture. These parameters were basically extracted from

differences in light transmission between neighboring im(D) age points within the nuclear contour. Two classes of

parameters were calculated:(E) (a) a class derived from differences in gray level between

pairs of image points along the scan lines in the nucleustaken in steps of 1, 2, etc., to 10 steps of points. The largerthe local variation in light transmission the more frequencies of differences for a given step length will be shiftedtowards larger values for those differences. The first moment of the frequencies after normalization according tonuclear size are used as parameters and are called DIFF1,DIFF2 DIFF1O depending on step length; (b) a classderived from a transition probability matrix over the nucleuspreviously used by Haralick et a!. (12) and Pressman (19).To facilitate processing we reduced the nuclear gray-levelrange to 8 gray levels. These parameters measure theaverage coarseness of chromatin structure in the nucleusfrom an 8 x 8 transition probability matrix. A set of 4parameters were calculated for each nucleus. The matrix

(F) had the following appearance:

p(lIlp(112 p(118p(211p(212 p(218i

p(811)P(812) P(818)

in whichp(ilj@is definedas the probabilityof gray-leveljoccurring after gray-bevel i in the nucleus when movingalong the scan lines within the nucleus. The 4 parameterswere:

ASM = ,@,@ p(i@f)2 (K)

CONTRAST= fl@,@ p(ilJ)] (L)

DIFMOMENT=@@ (i@ j)2.p(ilj) (M)

CANCER RESEARCH VOL. 384690


Computerized Nuclear Morphometry

@@[p(iIj _ m@COEFVAR= 64 m

p(iI/)m 64

Description of the TumorCell Population

To describe the tumor based on the nuclear measurements, we approximated the sample of multidimensionalrandom vectors that resulted from the measurements ofindividual nuclei by analytical probability distributions. Theparameters of the distributions were used as cell populationdescriptors. In this way the tumor cell population wasrepresented by a mean vector and a covariance matrix. Thecell population descriptors can be presented in the following way. Let@ be theJth descriptor,J = 1 . . . m, measuredfrom the ith cell, i = 1 . . . n, in the sample from a specifictumor. The means of the descriptors will then be

D@= 11,@,d,@

D@can be used as a tumor descriptor. The covariancematrix

Cj.k@@ @,(d,@D@(d,@0k)

can be used as another descriptor of the tumor.

OptimizationProcedure

Many methods can and will be used to evaluate theimportance of cancer cell population descriptors. For instance, the descriptors can be evaluated on the basis ofpatient survival, metastatic spread at the time of diagnosis,epidemiobogical characteristics, etc. In this basic study thedescriptors were used to model the degree of atypia basedon the cytological judgment.

In this study the nuclear atypia of a tumor cell populationlevel was described by a linear mathematical model:

CNM = nuclear population atypia =@ wD.

D is a population descriptor and w1is the weight assignedto that particular parameter. By choosing suitable weightsWi, the nuclear population atypia measure can be altered(under the limitation of linearity) to correspond to variouspatient parameters.

For adjustment of weighting, photographs of recordedcell populations were ranked by increasing atypia according to cytological criteria. The weights w were determinedfrom the atypia ranking by using multiple linear regressionanalysis performed stepwise with one complete regressionper step. At each step the one parameter that caused thelargest reduction in the error sum of squares was added. Acombination of m(AREA), m(AVLEVEL), m(ENTROPY),m(BDYVAR), m(COEFVAR), and m(DIF1) was first included

with the corresponding sample correlation coefficient between computer and visual assessment of nuclear popula

(N tion atypia.HistopathobogicalMethods

When the tumor was divided, a slice about 3 mm thickwas taken through the greatest diameter of the tumor, thusensuring that the central portion of the tumor was includedas well as the borders. This slice was fixed immediately inCarnoy's solution. The remainder of the tumor, surroundingtissues, and axillary tissues were fixed in 4% neutral formaldehyde for further study. The Carnoy's-fixed portion wassectioned and stained with the following: periodic acidSchiff stain with and without prior diastase digestion (mucinand glycogen); Voerhoff's orcein-iron hematoxylin (elasticfibers); Azan-Heidenhain's stain (connective tissue); methylgreen-pyronin (mast and plasma cells), and Alcian greenstain (keratin and mucus). In addition, formalin-fixed aswell as Carnoy's-fixed tissue was sectioned and stainedwith hematoxylin-eosin or van Gieson's stain and used formorphological diagnosis. Similar staining studies weremade of lymphatic metastases when warranted. Histobogicabbyconfirmed lymphatic metastases were found in 53 of

(0) the patients.Gradingand Classification

Each tumor was graded by the WHO (5, 21) and Black (4)systems. In addition, tumors were classified by the Ackerman (1), WHO (21), and the Armed Forces Institute ofPathology (25) classifications. These gradings and classifications were performed with the use of supplementarystains and histochemical methods and represent a moredetailed examination than that usually given breast cancercases. To arrive at a grade or classification, 2 pathologistsexamined the slides from each case and then reexaminedthe same slides after a passage of 6 months. Thus eachcase was graded or classified 4 different times. The variablereproducibility of the grading and classification systemsrequired that a system be developed for arriving at anapproximated final score. If 3 or 4 of the 4 readings of thetumor were in agreement, the tumor received the designation of these judgments. In cases in which only 2 readingsof 4 were in agreement, the final designation for thatspecimen was made by the pathologist who showed the

(Q) highestrateofreproducibilityofsubjectivegradingfor thatparticular system.

The tumors were also categorized to indicate the degreeof differentiation of the tumor tissue by dividing the tumorsinto low, middle, and highly differentiated.

Statistical Methods

Statistical comparisons were made between CNM valuesand classification and gradation variables. CNM is an interval scale variable. In determination of the linearity of therelationship between CNM and some other interval variable,the production moment correlation coefficient and its corresponding 2-tailed p value were used as a test of linearindependence (15). When ordinal variables or interval vanables that are not normally distributed were compared withCNM values, Spearman's rank correlation coefficient was

where

(P)

DECEMBER1978 4691


AREA5.555(454-0.256(451

ENTROPY5.455(45k5.155(451

DIFFI8.755(451-

--8.185(451

DIFF85.785([email protected](455

DIAGlIOII 8.155(482

8.555(485EXTDIC5.355([email protected](481Ai4@EV

5.555(452--

-@nA@fl,Lrc-8.188(451DIFF45.785(4515.588(455

kcONTRAST5.155(451

-

B. Stenkvist et a!.

used and , if the independence hypothesis was to be tested,its corresponding 2-tailed p value was also analyzed (14).

For comparisons with nominal variables such as theclassification systems, correlation coefficients for nominalscales such as Tschuprow's coefficient (26, 27) on Cramer'sV (7) were used. In some instances, correlation coefficientsmay have questionable reliability, and in those instancesthe most frequent categories in the nominal variable weretaken out and the t test for independent samples (6) or theMann-Whitney U test (23) were used to determine whetherthe CNM values for tumors in one category are significantlydifferent from those in another category.

RESULTSFig. 1 shows the quality of the cytological preparations

stainedwith chromealumgallocyanin.A comparisonbetween a high-resolution light photomicrograph of a cellnucleus and a computer image of the same nucleus isshown in Fig. 2. The image stored in the computer is at ahigher geometrical and photometric resolution than thatshown by the plotter, and for that reason the image issomewhat coarser than that recorded on film.

A binary map was made of the images of the cellsscanned from each tumor. One such map is shown in Fig.3. A frequency distribution of CNM factors for these cells isshown in Chart 2.

The role played by the cell sampling method was studiedby comparing a population of cells collected from the sametumor both by taking imprints of the cut surface of thetumor and by using fine-needle aspiration biopsy. Whenplotted against each other, the 2 populations of cells gave

Chart2. Histogramsof the distribution of differentparameters of the nuclei in Fig. 3 showing the density,shape, and texture parameters.

CNM values with a correlation coefficient of 0.79 (Chart 3).Chart 4 demonstrates a scatter plot of the result of

multiple-site sampling from the same tumor by repeatedfine-needle aspiration biopsy. The correlation coefficientfor these values is 0.86, obtained by plotting 2 differentpopulations against each other. Because of the good conrelation obtained between the CNM values of these apparently geometricallymonocbonabcells, we used this set ofdata to test some of the individual descriptors that make up

659 550

r@iPGRADE

@9150181

Chart3. Scatterplot illustratingthe correlationbetweenCNMvaluesforcell populationstakenfromthe sametumor byfine-needleaspirationbiopsy(DATGRADE)and by imprints (IMPGRADE) from the cut surface ofthe tumor.The coefficient of variation for the reproducibility of CNM is 0.79.

/L

5.855(455 @CEF@i S. 1@4S2

5.555(455 BDYV55II 5.355(453

CANCER RESEARCH VOL. 384692


ProductSpearmoman'smentrankcome

correlaNo.lationCorre tionCorreofcoeffi spend coeffi spend

pairscientingpcientingpSubjectivejudg1370.4570.0000.4400.000ment

of mitosesDegree

of differ1370.0270.7580.0410.632entiationGrade

according1370.4760.0000.4630.000toBlackGrade

according1370.4330.0000.4220.000toWHO


731.150REPGRADE

Table 1Correlationswith CNMfor 4 variables(including 2-tailed valuesof

the independence hypothesis) used as positive and negativestatistical controls

180.850@52.

Chart 4. 5catter plot showing the correlation between CNM values forcells taken from different parts of the same tumor. One population value onthe horizontal axis corresponds to the value on the vertical axis that wasobtained by measuring another population from a different part of the tumor.Correlation coefficient, 0.86.

9227.100

RM

FFI

1230.9001578 .

Chart 5. Average DIFF4 values measured twice from each of 14 tumors,where the first determination is plotted on the horizontal axis (MDIFF4) andthe second is plotted on the vertical axis (RMDIFF4) as a scatter diagram.

the CNM value. Chart 5 shows the comparison of one suchhighly abstract measure based on the frequencies of difference in gray levels, average DIFF4 values, with a correlationcoefficient of 0.83. Other such comparisons gave correlation coefficients that varied between 0.75 and 0.86.

Two additional measures were added for comparisonwith CNM values, mitoses and subjective assessments oftumor differentiation. One of these, mitoses, was intendedas a positive control of CNM and was expected to becorrelated with a high CNM value, while the other, differentiation degree, was expected to be independent of CNM.The results of these comparisons are seen in Table 1.

Another comparison between a subjective grading systemand CNM was made with the Black grading system (Chart 6A). The product moment correlation coefficient between

the Black grading and CNM was 0.47 and between the WHOgradingand CNM was 0.42(Table1).

A comparison was also made between final pathologicaldiagnoses based on the different classification systems andCNM (Chart 6, B to 0).

DISCUSSION

The texture of the nuclear image is an important component in the histopathological-cytopathological determination of cancer. For this purpose cells are fixed and stainedin such a way that a reproducible pattern is obtained andthis pattern is then judged subjectively by the eye of thepathologist. The chrome alum gabbocyanin used here isgenerally believed to form (specific) stoichiometnic complexes with both RNA and DNA. Thus CNM is most probablya measure of nucleic acid content (i.e., cell pboidy, transcniption rate) and distribution (i.e. , euchromatin, heterochromatin, RNA distribution) within the nucleus. Based onits specificity, stoichiometry and intensity, chrome alumgalbocyanin appears to be a better choice for use with CNMthan does a non-stoichiometnic, classical staining methodsuch as hematoxylin-eosin.

In a recent paper by Nicolini et a!. (16), it was reportedthat the texture of Feulgen-stained Wl-38 nuclei changedwhen cell growth was stimulated by addition of serum tothe medium, although the DNA content of the cells remained the same. These textural changes were apparentlyrebated to the cell cycle: cells in Go or G2 arrest had adifferent nuclear DNA morphology than did cells in otherportions of the cell cycle. They also found that chromatindispersion is followed by a modulation of the nuclear shapeto give an irregular, less circular shape to the nuclearborder. This report, coupled with the positive correlationbetween CNM value and number of observable mitoses in atumor (Table 1), is reason to believe that another factorinfluencing CNM values is the growth activity of the cellpopulation. The possible correlation between growth activity of a tumor cell population and CNM values should befurther explored because of its implications for choice andplanning of chemotherapy. That is, many chemotherapeutic

DECEMBER1978 4693


B. Stenkvist et a!.

Gradin(B1ackN: 31

N: 2

6A

CLASSIFICATIONS AFTER ARLED 0ORCESINSTITUTE OF PATHOLOGY

LOBULAR CARc:N0VA Ii!:V

IINVASIVE)

CGLLOID CARCIGOMA

MEDULLARY CARCIr,O.G

CARCINOMA WITH F:BRos:s

COMEDOCARC INOVA

PAPILLARY CARCINOMA

CLASSIFICATION ACCORDING TG INC

LOBULAR CARCINOMA C:111:E

MUCOUS CARCINOMA C:I11:o

6C

84@

I1:B:3

II:s:2

N: 34

CLASSIFICATION ACCORDING TO

ACKERMAN

HIGHLY METASTASIZING I@

INTRADUCTAL CARCINOMA 111:2WITH STROMAL INVASION

ADENOCARCINOMA 111:1

WELL-DIFF. ADENOCARCINOMA 1:3

175 CNMVALUE

175 C4M VALUE

CNN value 844

6B

844

6D

PAPIlLARY CARCINOMA C:111:B

MEDALLARY CARCINOMA 11:2MEDALLARY CARCINOMA C:III:@

INVASIVE CARCINOCA C:1111:1MUCINOUS CARCINOMA

844

5

175 CNMVALUEChart 6. Distribution of CNM values. A , grading system advocated by Black; b , Ackerman classification system ; c , tumors classified by the AFIP criteria; d,

tumors classified according to the WHO system. 01FF, differentiated.

agents are effective only in S phase and, if a major portionof the cells are in a resting phase, these antitumor agentswill not be effective. Prediction by CNM of this effect (orback of effect) would be a valuable clinical tool in cancertreatment.

Low reproducibility of subjective grading and classification systems as found in this study has been reportedpreviously. Cutler et a!. (8) studied the reproducibility ofBlack's grading system and achieved results similar to ours,although a thorough statistical study of the results was notconducted. Schiodt (22), in an analysis of the gradingsystem of Bloom and Richardson (5), reported a lowerreproducibility than did Bloom and Richardson.

A common assumption in a situation in which poorresults are obtained in subjective judgments is that thoseobservers who do not succeed are in some way less competent than the individual who devised the system. In thiscase, the pathologists making the subjective judgmentsrepresent at least an average standard of competence inbreast cancer diagnosis. They used a number of additionalhistopathobogicab methods and certainly devoted more timethan is usual to making diagnostic decisions but failed toachieve consistent reproducibility (24). We believe thatCNM, with its great reproducibility, can therefore serve as aremarkable aid in helping the pathologist to a decision andthus form a new part of the armamentanium of cancerdiagnosis.

CNM is capable of providing both a continuous intervalscale variable and a group of independent descriptors.These allow statistical comparison between other charactenistics of the patients' disease such as hormonal status,estrogen receptors, thyroid function, and the completebattery of chemical findings with an objective measure ofthe tumor. Such a study has now been completed; theresults will be reported elsewhere.

As with any system with possible patient diagnosis application, the question of throughput and cost effectivenessshould be addressed. To scan one cell for CNM requiresabout 4 mm depending on the quality of the slide and theadeptness of the operator. Thus the instrument is capableof processing one patient per day in its current configuration. If justified by clinical applications, the throughput ofthe system can be increased at least 50-fold for routinescanning. The greatest amount of time is currently spent inlocating cells suitable for study. Most of the operations canbe performed by competent technical assistants providedtheir judgments of cytopathology are examined by a pathologist. Costs for such an instrument can probably not bejustified at present except as a research procedure, andfurther estimates of its usefulness must await correlation oflong-term clinical observations with CNM values.

The Swedish population registry and the availability ofcomplete hospital records along with the exhaustive, population-based cohort epidemiological study in this project

4694 CANCERRESEARCHVOL. 38



ensure that clinical course and prognosis is certain to beobtained for all of the patients, as it has been for the last 2years.

This study has demonstrated that individual cells obtamed from human tumors by fine-needle aspiration biopsyare suitable for high-resolution image analysis. CNMthrough application of weighting optimization also allowsapplication of cytopathobogical data to the entire spectrumof clinical data such as chemical, radiological, or epidemiobogical information about patients.

Application of CNM to other solid tumors will be a majorpart of the future for this apparatus. Not only breast tumorsbut also other epithelial tumors (with only one or a fewmajor cell types) such as those of the prostate and thyroid,as well as lymphomas, are suitable for study. We believethat CNM, with its great reproducibility, can be a remarkablyuseful aid in helping the pathologist to more exact gradingof the malignancy of a tumor.

REFERENCES

1. Ackerman, L. v. , and del Regato, J. A. Cancer, Diagnosis Treatment andPrognosis, Ed. 4. St. Louis, Mo.: The C. I.Mosby Co., 1970.

2. Bengtsson, E. On the Design of Systems for Computer Aided Analysis ofMicroscopic Images. Doctoral Thesis, Uppsala University, Uppsala,Sweden, 1977.

3. Bengtsson, E., Eriksson, 0., Holmquist, J., and Stenkvist, B. A SoftwareSystem to Record and Analyze Digitized Cell Images. Computer Progr.Biomed.,7:233-246,1977.

4. Black, M. M., and Spear, F. D. Nuclear Structure in Cancer Tissues.Surg. Gynecol. Obstet., 105: 97-102, 1957.

5. Bloom, H. J. G., and Richardson, W. W. Histologic Grading and Prognosis in Breast Cancer. A Study of 1409 Cases of Which 359 Have BeenFollowed for 15 Years. Brit. J. Cancer, 11: 359-377, 1957.

6. Brownlee, K. A. Statistical Theory and Methodology in Science andEngineering, Ed. 2. New York: John Wiley & Sons, Inc., 1967.

7. Cramer, H. Mathematical Methods of Statistics. Uppsala, Sweden:Almqvist & Wiksell, 1945.

8. Cutler, S. J., Black, M. M., Friedell, G. H., vidone, A. A., and Goldenberg, I. S. Prognostic Factors in Cancer of the Female Breast. II.

Reproducibility of Histopathologic Classification. Cancer, 19: 75-82,1966.

9. Einarson, L. Gallocyanin-Chromlackreaktion als Grundlage eines ZellSquivalentbildes. Laeknabladid, 20: 142-153, 1934.

10. Einarson, L. On the Theory of Gallocyanin-chromalum Staining and ItsApplication for Qualitative Estimation of Basophilia. A Selective Stainingof Exquisite Progressivity. Acta Pathol. Scand., 27: 82-92, 1951.

11. Granlund, G. H. Fourier Pre-processing for Handprint Character Recognition. IEEE Trans. Computers, C-21: 195-201 , 1972.

12. Haralick, A. M. , Shanmugam, K., and Dinstein, I. Textural Features forImage Classification. IEEE Trans. Systems Man Cybernetics, SMC-3:610-622,1973.

13. Holmquist, J. On Analytic Methods and Software Design for ComputerProcessing of Digitized Microscopic Cell Images. Doctoral thesis, Uppsala University, Uppsala, Sweden, 1977.

14. Kendall, M. G. Rank Correlation Methods. London: Charles Griffin, andCo., Ltd., 1948.

15. Kendall, M. G., and Stuart, L. A. The Advanced Theory of Statistics, Vol.2. London: Charles Griffin, and Co., Ltd., 1967.

16. Nicolini, C., Giaretti, W., De Saive, C., and Kendall, F. The G0-G1Transition of WI 38 Cells. II. Geometric and Densitometric TextureAnalysis.Exptl.Cell Res.,106:119-125,1977.

17. Norberg, A. OSIRIS-Electronics for Computer Control of an ImageScanner with a Vibrating Prism and a Diode-Array. Doctoral Thesis,Physics IV. Royal Institute of Technology, Stockholm, 1976.

18. Nordin, J. OSIRI5The Opto-mechanical Part of a Scanner Using aDiode-Array to Digitize Microscope Images. Doctoral Thesis, Physics IV,Royal Institute of Technology, Stockholm, 1976.

19. Pressman, N. J. Markovian Analysis of Cervical Cell Images. J. Histochem.Cytochem.,24: 138-149,1976.

20. Prewitt, J. M. S., and Mendelsohn, M. L. The Analysis of Cell Images.Ann. N. Y. Acad. Sci., 128: 1035-1053, 1966.

21. Scarff, R. W., and Torloni, H. Histological Typing of Breast Tumours.International Histological Classification of Tumours No. 2. Geneva:WHO, 1968.

22. Schiodt, T. Breast Carcinoma. A Histologic and Prognostic Study of 650Followed-up Cases. Copenhagen: Ejnar Munksgaard, 1966.

23. Siegel, 5. Nonparametric Statistics for the Behavioral Sciences. NewYork: McGraw-Hill,BookCo., 1956.

24. Stenkvist, B., Westman-Naeser, 5., Vegelius, J., Holmquist, J., Nordin,B., Bengtsson, E., and Eriksson, 0. Analysis of the Reproducibility ofSomeSubjectiveGradingSystemfor BreastCarcinoma.J. Clin. Pathol.,in press, 1979.

25. Stewart, F. w. Tumors of the Breast Atlas of Tumor Pathology, Sect. Ix.Washington,D.C.:ArmedForcesInstituteof Pathology,1950.

26. Tschuprow, A. A. Grundbegriffe und Grundprobleme der Korrelationstheorie. Stuttgart: B. G. Teubner, Verlagsges. MbH, 1925.

27. Tschuprow, A. A. Principles of the Mathematical Theory of Correlation.New York: William Modge, 1939.

DECEMBER1978 4695


B. Stenkvist et a!.

@@1

*

IL

@1'@S

:

.

.@ @...:.@

4696 CANCER RESEARCH VOL. 38

a..

@...Fig. 1. Photomicrograph (initial magnification, x 500) of a typical cell population from a fine-needle aspiration biopsy. Arrows, typical cells that would be

selected for measuring.Fig. 2. Photomicrograph (initial magnification, x 1000 at N. A. 1.30) of a typical cancer cell nucleus obtained from fine-needle aspiration biopsy and

stained with gallocyanin-chrome alum and the image of the same cell obtained by the computer. Resolution of the computer image is deceptively lowbecause of limitations of the plotter resolution. There is some geometrical distortion due probably to a display artifact.


I_@._____

... I


DECEMBER 1978 4697

1 . 0.

Fig. 3. Binary plots of nuclei from 40cancer cells obtained from a fine-needlebiopsy of a human breast cancer.

S S.

S S . S 0

.11 I 5


1978;38:4688-4697. Cancer Res

Bjrn Stenkvist, Sighild Westman-Naeser, Jan Holmquist, et al.

Characterizing Human Cancer Cell PopulationsComputerized Nuclear Morphometry as an Objective Method for

Updated version

http://cancerres.aacrjournals.org/content/38/12/4688Access the most recent version of this article at:

E-mail alerts related to this article or journal.Sign up to receive free email-alerts

SubscriptionsReprints and

[email protected] atTo order reprints of this article or to subscribe to the journal, contact the AACR Publications

Permissions

[email protected] atTo request permission to re-use all or part of this article, contact the AACR Publications


computerized nuclear morphometry as an objective method for characterizing human cancer cell...

Documents

cancer research

nucleiof breast cancer

breast cancer task force

karolinska institute

tumor weresampled

uppsala university

grading systems

thenew method