j expert systemsupport bayesianbeliefgist entered evidence into the networkin the formoflikelihood...

J Clin Pathol 1994;47:329-336

Expert system support using Bayesian beliefnetworks in the diagnosis of fine needle aspirationbiopsy specimens of the breast

P W Hamilton, N Anderson, P H Bartels, D Thompson

AbstractAim-To develop an expert system modelfor the diagnosis of fine needle aspirationcytology (FNAC) ofthe breast.Methods-Knowledge and uncertaintywere represented in the form of aBayesian belief network which permittedthe combination of diagnostic evidence ina cumulative manner and provided afinal probability for the possible diagnos-tic outcomes. The network comprised 10cytological features (evidence nodes),each independently linked to the diagno-sis (decision node) by a conditional prob-ability matrix. The system was designedto be interactive in that the cytopatholo-gist entered evidence into the network inthe form of likelihood ratios for the out-comes at each evidence node.Results-The efficiency of the networkwas tested on a series of 40 breast FNACspecimens. The highest diagnostic proba-bility provided by the network agreedwith the cytopathologists' diagnosis in100% of cases for the assessment of dis-crete, benign, and malignant aspirates.Atypical probably benign cases weregiven probabilities in favour of a benigndiagnosis. Suspicious cases tended tohave similar probabilities for both diag-nostic outcomes and so, correctly, couldnot be assigned as benign or malignant. Acloser examination of cumulative beliefgraphs for the diagnostic sequence ofeach case provided insight into the diag-nostic process, and quantitative datawhich improved the identification of sus-picious cases.Conclusion-The further development ofsuch a system will have three importantroles in breast cytodiagnosis: (1) to aidthe cytologist in making a more consis-tent and objective diagnosis; (2) to pro-vide a teaching tool on breast cytologicaldiagnosis for the non-expert; and (3) it isthe first stage in the development of asystem capable of automated diagnosisthrough the use ofexpert system machinevision.

(3 Clin Pathol 1994;47:329-336)

Fine needle aspiration cytology (FNAC) hasbeen established as a rapid, safe, and costeffective method of diagnosis in breast dis-ease. The success of the technique reliesstrongly on the ability of the cytologist to

identify and characterise cytological changesin the prepared aspirate. This presents anumber of problems. Diagnosis is largelybased on visual criteria which are subjectiveand can be misleading in certain instances.The number of visual clues which need to beassessed and the number of options availableimpose difficulties for assimilating all the rele-vant diagnostic information in a consistentand reproducible manner. This is also true ofother areas of histological and cytologicaldiagnosis. 1

Expert systems are computer programs thatare designed to store, access, and processknowledge about a particular domain.2 Theycan therefore provide the perfect frameworkfor storing cytological diagnostic knowledge ina logical, consistent, and reproducible mannerand have substantial potential in providingcytopathologists with a means of making amore accurate and consistent diagnosis.

Decision making in cytopathology (as inother domains) involves the consideration andcombination of numerous pieces of evidence.When the diagnostic evidence is clear-cut, theuncertainty involved in reaching a decision islow. However, most of the knowledge in diag-nostic cytopathology is in the form of images,concepts, and descriptive terminology andbecause of this, uncertainty in the decisionsleading to diagnosis and in the final diagnosis isinherent. It is for this very reason that expertsare required to make difficult diagnostic deci-sions. The development of computer-basedexpert systems requires careful considerationas to how uncertainty is handled and howknowledge is to be represented within thedesign3 4

Bayesian belief networks provide a power-ful means of representing knowledge: evi-dence is combined in a cumulative manner toprovide a measure of certainty in the finaldecision. Their structure is comprised ofNODES connected by LINKS (fig 1A),where the nodes constitute probabilistic vari-ables-nuclear pleomorphism-with possibleoutcomes-none, mild, moderate, severe-and links represent the relation between twonodes and are quantified by a conditionalprobability (CP) matrix. The CP matrixexpresses the relation between the possibleoutcomes of the descendent child node withthe outcomes of the parent node. When thedescendent node represents evidence, such as acytological feature, and the parent node is thediagnosis as in the current study, the CPmatrix expresses the probability of finding afeature outcome (such as pleomorphism:

Department ofPathology, TheQueen's University ofBelfast, NorthernIrelandPW HamiltonN AndersonOptical SciencesCentre, University ofArizona, Tucson,Arizona, USAP H BartelsD ThompsonCorrespondence to:Dr Peter W HamiltonQuantitative PathologyLaboratory, Institute ofPathology, The Queen'sUniversity of Belfast,Grosvenor Road, BelfastBT12 6BL, N Ireland.Accepted for publication27 October 1993

329

on May 7, 2020 by guest. P

rotected by copyright.http://jcp.bm

j.com/

J Clin P

athol: first published as 10.1136/jcp.47.4.329 on 1 April 1994. D

ownloaded from

http://jcp.bmj.com/

Hamilton, Anderson, Bartels, Thompson

eDiagnosisOutcomes:

Benign Malignant (

CP | inkMatrix I

Outcomes-: odNone Mild Moderate Severe (evidence

Conditional probability matrix

None M

Diagnosis Benign 0-5 CMalignant 01 0

Node(decision)

BPleomorphismAild Moderate Severe0-4 0-09 0-0102 03 0.4

update of all the prior probabilities and beliefsat the other evidence nodes (fig 2). At thedecision node, accumulating evidence willresult in changes in the outcome belief untilthe belief in a particular outcome exceeds adefined threshold, sufficient to constitute adiagnostic decision.A detailed description of the algorithmic

mechanism of network initialisation and beliefpropagation in Bayesian belief networks isgiven by Bartels, Thompson, and Weber.4Bayesian belief networks have been applied inthe diagnosis of prostate lesions56 and theirtheoretical application in quantitative histo-pathology has been extensively researched byBartels et al.7

Figure 1 (A) Diagram of basic Bayesian belief network unit comprising a decision node(DIAGNOSIS) linked to an evidence node (PLEOMORPHISM) via a conditionalprobability (CP) matrix. (B) This shows an enlargement of the CP matrix link whichconveys the probability ofpleomorphism outcomes being present given the diagnosis.

Table 1 List of diagnosticfeatures and their possibleoutcomes used inconstruction ofBayesianbelief network

Bare nuclei:numerous, moderate,few/absent

Cellularity:low, moderate, high

Cohesion:strong, moderate, poor

Pleomorphism:none, mild, moderate,severe

Cell arrangement:even/flat, irregular/overlap

Nuclear size:small, medium, large

Nucleoli:none, single, multiple

Intracytoplasmic lumina:present, absent

Aprocrine cells:present, absent

Mucinous background:present, absent

mild) given a diagnostic outcome (such as

benign) (fig iB).The belief in the various outcomes is also

stored at each node. This is expressed as a

probability. For the decision node, this isdefined by the user. For the evidence nodes,the outcome probabilities are defined by a

network initialisation process and are largelydependent on the values held in the CPmatrices.'The system reported here is designed to be

interactive in that the cytopathologist exam-

ines the cytological features of a case andenters evidence into the network in the formof relative likelihood ratios. These convey thelikelihood that each of the possible featureoutcomes are present in the sample underinvestigation (see table 5 for examples).Values of zero are not used as this presentscomputational problems. The value for eachoutcome is an independent estimate and thetotal does not have to equal 100. They there-fore define in numerical form the subjectiveimpression of the cytologist.When evidence is submitted to an evidence

node, the belief in the outcomes of that nodeare updated. This belief is then propagated, inour very simple network, upwards to the diag-nostic decision node, where the belief in thediagnostic outcomes (benign and malignant)are updated. This update, in turn, triggers an

Evidence

Figure 2 Propagation of belief in a Bayesian belief network. * = nodes, = links.Arrows pointing up: update of likelihoods and beliefs. Arrows pointing down: update ofprior probabilities and beliefs. This illustration shows a second level of descendent nodes atwhich the evidence was entered. The network in the current study only possessed a singlelayer of evidence nodes.

MethodsBAYESIAN BELIEF NETWORK DESIGNThe Bayesian belief network was constructedusing software developed in The OpticalSciences Centre, University of Arizona, USA.The program was written in C and has agraphical interface which permits the inter-active definition of nodes, links, and CPmatrices. Algorithmic implementation fol-lowed that designed by Morawski89 from theconcepts originally developed by Pearl.'0A set of 10 diagnostic clues were defined

which provide evidence in the diagnosis ofbreast FNAC. These are listed in table 1together with their possible outcomes. Notethat these features were chosen for the assess-ment of aspirates stained using Hema-Gurr(BDH Chemicals), the method used in ourlaboratory where a rapid diagnostic service isprovided at outpatient clinics.The belief network topology was in the

form of an open tree with a single level of evi-dence nodes (fig 3). This simple design con-sisted of a single decision node representingthe DIAGNOSIS with its possible outcomes(BENIGN and MALIGNANT), and the 10diagnostic features as evidence nodes directlylinked to the diagnostic decision node.Conditional probability matrices relating eachof the diagnostic clues and their outcomes todiagnosis (benign, malignant) were defined bya cytopathologist (NA) (table 2).

CASESThe performance of the belief network wastested using a series of 40 cases of breastaspirates retrieved from file at the RoyalVictoria Hospital, Belfast, all of which hadconfirmatory tissue biopsy diagnoses. Basedon the cytology report, these comprised 16benign, 11 malignant, eight suspicious, andfive atypical probably benign cases using thecriteria of the NHS Breast ScreeningProgramme" (table 3). Each case wasassessed cytologically for all of the featureslisted in table 1 and relative likelihood ratiosrecorded and entered as evidence into eachnode of the network. After each piece of evi-dence was submitted the calculated probabili-ties for the diagnostic outcomes wererecorded. After all the evidence nodes weresampled for a single case, the final diagnostic

330



j.com/

J Clin P


ownloaded from

http://jcp.bmj.com/

Expert system for FNA diagnosis of breast

Figure 3 Topology ofBayesian belief network forthe diagnosis of breastFNAC. The decision nodeis DIAGNOSIS and theevidence nodes are thecytologicalfeatures. Theoutcomes for each node arealso shown.

Diagnosis:Benign Malignant

Bare nuclei Mucinous backgroundNumerous Moderate Few/ A Present Absent

Celularity Apocrine cellsLowModerate High Present Absent

CohesionderateHig Intracytoplasmic luminaStrong Moderate Poor Pleomorphism S umNudoleo Present Absent

None Mild Moderate Severe SalMcimLre None Single MultipleColl armngement

Even/lfat Irregular overlap

probability was compared with the originaldiagnosis given by the cytopathologist.

Table 2 Conditional probability matrices relatingprobability offindingfeature outcome, given the diagnosis(these sum to 1 0 in the horizontal)

Bare nucleiNumerous Moderate

Benign 0-55 0-40Malignant 0-01 0-10

CellulaityLow Moderate


CohesionStrong Moderate


PleomorphismNone Mild


Fewlabsent0-050-89

High0-100-85

Poor0-050-50

Moderate Severe0-09 0-010-30 0-40

Cell arrangementEven/flat Irregularloverlap


Nuclear sizeSmall Medium


NucleoliNone Single


Large0-010-40

Mulnple0-010-40

Intracytoplasmic luminaPresent Absent


Apocrine cellsPresent Absent


Mucinous backgroundPresent Absent


Table 3 Diagnostic categories and summary of diagnosticcriteria used in this study and defined by NHS BreastScreening Programme"

Benign Indicates an adequate sample showing noevidence of malignancy

Atypical The aspirate can have all theprobably characteristics of a benign aspirate, butbenign certain features not commonly seen in

benign aspirates may be present

Suspicious of The material is not diagnostic ofmalignancy malignancy but may show some

malignant features without overtmalignant cells being present

Malignant Indicates an adequate sample containingcells characteristic of carcinoma orother malignancy

NETWORK PERFORMANCEThe value of belief in the diagnostic out-comes, as provided by the network, was usedto determine the ability of the system to diag-nose correctly breast aspirates. Belief thresh-olds were defined so that cases could beallocated into groups based on their networkderived diagnostic probability and comparedwith the original diagnosis. If a case had aprobability of >0-90 it was recorded as aclear-cut BENIGN or a MALIGNANT caseby the network. Cases which had a belief inbenign of <0 90 but >0 50 were calledEQUIVOCAL (B). Cases which had a beliefin malignant of <0 90 but >0 50 were desig-nated EQUIVOCAL (M).

ADDITIONAL DIAGNOSTIC MEASUREMENTSCumulative diagnostic probability graphsProbability graphs were drawn for each case,plotting the cumulative diagnostic probabilityafter each piece of evidence was submitted (fig4). The evidence nodes were examined in anorder considered by the cytologist to be ofdecreasing importance to the final diagnosis.Graph characteristics could be measured insimple terms: the number of peaks (P-score),troughs (T-score), and intersects with the 0-5,0 5 threshold (C-score) were determined foreach case. These measure the extent of con-flicting information across the range of diag-nostic clues.

0E00

.0

-

m

Diagnostic cluesFigure 4 Cumulative probability graph for a benignfibroadenoma. Notice that it is the presence of bare nucleiwhich has the strongest initial effect in keeping the line inbenign halfof the chart.

331



j.com/

J Clin P


ownloaded from

http://jcp.bmj.com/


Table 4 Binary code for absolute benign cases andfor malignant cases

CellBare nuclei arrangement

Cellulaity Cohesion PleomorphismFewl Evenl Irreg.

Numerous Moderate Absent Low Moderate High Strong Moderate Poor None Mild Moderate Severe flat overlap

Absolute benign cases:1 0 0 1 0 0 1 0 0 1 0 0 0 1 0

Malignant cases:0 0 1 0 0 1 0 0 1 0 0 1 0 0 1

Binary coding and distance measuresThis required defining a binary code for ahypothetical "absolute" benign case. For eachdiagnostic feature, the most typical outcomefor a benign case was given a value of 1, andthe other outcomes a value of 0 (table 4).Therefore, the binary code for benignity is:100l001001000l0100l000l01 against whichthe codes for real cases were compared.Apocrine metaplasia was omitted from thebinary scoring system, because while its pres-ence is indicative of benignity, its absence hasno diagnostic power. Each case in the serieswas allocated a similar code based on themost prevalent outcome of each feature (table4).

Deviation away from the absolute normalbinary code was quantified using two distancemeasures. Firstly, the Hamming distancemeasures the number of digits which changeto produce the new code.12 The greater thedeviation away from the normal blueprint, thelarger will be the Hamming distance.A second measure was used where the

sequences of diagnostic clue values were con-sidered as a string, with substrings for eachclue. The total number of jumps necessary toproduce the new clue value for each clue inde-pendently were counted and summed. Thiswe termed the "diagnostic clue distance".These measures were made for each case inthe series as additional evidence for the diag-nostic classification of cases.

Table 5 Likelihood ratios for a benign case number 1 *

Bare nuclei:Numerous, moderate, few/absent0-30 0-90 0-30

Cellularity:Low, moderate, high0 75 0-65 0 10

Cohesion:Strong, moderate, poor0 90 0-60 0 05

Pleomorphism:None, mild, moderate, severe0-85 0-75 0-15 0-01

Cell arrangement:Even/flat, irregular/overlap0-80 0 10

Nuclear size:Small, medium, large0-95 0-25 0 01

Nucleoli:None, single, multiple090 050 010

Intracytoplasmic lumina:Present, absent0 10 0-95

Aprocrine cells:Present, absent0-01 0-99

Mucinous background:Present, absent0-10 0-98

*Enteing this evidence into the Bayesian network resulted in aBenign/Malignant probability of 1-00/0 00 (see table 6)

REPRODUCIBILITY OF RELATIVE LIKELIHOODRATIOSThe same series of cases were relabelled andpresented to the same cytologist who wasblind to the original diagnosis and to the pre-viously defined likelihood ratios. They werereassessed and new likelihood ratios for eachcase entered into the network. Reproducibilityof case classification based on the networkdiagnostic probability was assessed.

ResultsLikelihood ratios for each of the diagnosticfeatures were estimated for all 40 cases. Anexample of a typical result for a benign case isshown in table 5.The diagnostic probability results provided

by the Bayesian belief network for all 40 casesare shown in table 6.

Table 6 List of cases with resulting diagnostic probabilityprovided by Bayesian beliefnetwork

Benign

Malignant

Atypical/benign }

Suspicious

Belief

Benign Malignant

1-00 0.000 59 0-411-00 0.00093 0071-00 0-000-99 0-010-99 0-011-00 0.001-00 0-001-00 0-000-99 0-011-00 0-001-00 0.001-00 0-001-00 0.001-00 0-00

0-00 1-000-00 1-000-08 0-920-00 1-000-00 1-000-00 1-000-00 1-000-00 1-000-00 1-00043 0570-00 1.00

1-00 0-001-00 0-000-98 0-020-68 0-32095 005

I 0-910-080-120050-840-260*000 34

0090-920-880950-160-741-000-66

332



j.com/

J Clin P


ownloaded from

http://jcp.bmj.com/


Intracytoplasmic MucinousNuclear size Nucleoli lumina background

Small Medium Large None Single Multiple Present Absent Present Absent

Absolute benign cases::1 0 0 1 0 0 0 1 0 1

Malignant cases:0 1 0 0 1 0 1 0 0 1

BENIGN AND MALIGNANT CASESOf the 16 cases diagnosed by the cytologist asbenign, all were given probabilities in favourof a benign diagnosis by the network (table 6).One case, however, fell into the EQUIVO-CAL (B) category with a benign/malignantbelief of 0-59/0-41. Review of this caseshowed that it was a fibroadenoma, a lesionwhich can pose diagnostic difficulties in cytol-ogy. Evidence for benignity came from theprobability graph (fig 4) which showed thatthe probability line remained completelywithin the benign half of the graph. While theHamming distance for this case had no dis-criminatory power, the diagnostic clue dis-tance was lower in the fibroadenomacompared with the malignant cases (fig 5).None of the benign cases had a positive C-score (table 7).Of the 11 malignant cases, again all were

given probabilities in favour of a malignantdiagnosis by the network (table 6). A singlecase had a lower probability than the rest(benign 0-43 malignant 0-57) putting it intothe EQUIVOCAL (M) group. On review thiswas found to be an aspirate from an infiltrat-ing lobular carcinoma, again a lesion difficultto diagnose from cytology. Examination of theprobability curve (fig 6) clearly showed con-flicting evidence with a C-value of 2. Thenumber of evidence nodes which support abenign belief was four as opposed to sixdownward shifts supportive of malignancy.

This lobular carcinoma was the only malig-nant case to show a probability curve whichcut the 0 50:0 50 threshold line (table 7).

ATYPICAL PROBABLY BENIGN CASESAll of the atypical probably benign cases hadbenign probabilities and one of these fell intothe EQUIVOCAL (B) category. None of thebelief curve characteristics or binary distancemeasures could be used to discriminate thesefrom certain benign aspirates.

SUSPICIOUS CASESOf the eight suspicious cases, four showedindeterminate probabilities meaning that theycould not be directly classified as benign ormalignant. Of these four, three showed higherprobabilities for malignancy (EQUIVOCAL(M)). All of these three cases were confirmedto be malignant on biopsy. The single casehaving a network benign probability of 0-84was shown to be an infiltrating lobular carci-noma. Three cases were classified as malig-nant by the network, all of which wereconfirmed to be malignant on biopsy. A singlecase was given a benign probability (>090)by the network, but examination of the proba-bility curve showed it to have a C-score of 1, aT-score of 3 and a P-score of 3.

All but one of the suspicious cases had aprobability curve which cut the 0-5:0 5threshold on one or more occasions (table 7).

REPRODUCIBILITY OF RELATIVE LIKELIHOODRATIOSRe-examination of the same cases resulted inslightly different likelihood ratio vectors.Correct case classification based on the finalprobability, however, was not significantlyaffected. Benign and malignant cases weremuch more clearly defined on the secondoccasion, with all benign cases having a1-00/0-00 benign/malignant probability andall malignant cases showing a 0-00/1-00benign/malignant probability. As before,

Figure S Comparison ofbinary distance measuresfor benign and malignantcases. Thefibroadenoma(F) which showed non-definite probability valueswas categorised into thebenign range on the basisof the diagnostic cluedistance.

Benign cases4A

3

2

1 2 3 4 5 6- 7 8 9 1011121314151617181920Hamming distanceMalignant cases

IC

Hamming distance

Benign cases

F

1 2 3 4 5 6 7 8 91011121314151617181920Diagnostic clue distance

Malignant cases3-D

2-

1 2 3 4 5 6 7 8 91011121314151617181920Diagnostic clue distance

1 2 3 4 5 6 7 8 91011121314151617181920

333



j.com/

J Clin P


ownloaded from

http://jcp.bmj.com/


Table 7 Number of cases falling into C-score, P-score, and T-score intervals, as calculatedfrom cumulative probabilityplots

C-score P-score T-score

0 1 2 0 1 2 3 4 0 1 2 3 4

Benign (n = 16) 16 0 0 11 2 0 3 0 11 2 1 2 0Atypicalprobablybenign (n = 5) 5 0 0 3 0 0 2 0 3 0 1 1 0Suspicious (n = 8) 1 3 4 0 1 4 2 1 0 1 1 5 1Malignant (n = 11) 10 1 0 8 1 2 0 0 8 1 0 2 0

_*- X , | | |

_ /

,. . . ~~~~Nl. .

l l

so 0

~ ~

l .41

suspicious cases fell into the equivocal cate-gory with clear signs of conflicting evidence intheir cumulative probability curves.

DiscussionThe role of expert systems in the diagnosis ofbreast pathology has been considered before.Heathfield et al developed a rule based expertsystem for the diagnosis of breast FNACbased on a series of IF ... THEN rules whereuncertainty in individual rules and in the finaldecision was not assessed.'3 This approach is a

commonly used method of representingknowledge in expert system development, butit is questionable whether diagnostic expertisecan be reduced to a series of simple monotonicrules. The incorporation of certainty measuresin rule based systems, the methods used toupdate certainty and the handling of missingdata, can also be problematic.'0 14 Bayesianbelief networks, however, can provide a reli-able method for handling certainty which isbased on well established probability calculus.They can also operate to completion andreach a final diagnostic belief even when evi-dence is missing. Some of the problemsassociated with rule based knowledge repre-sentation were recognised in a later paper byHeathfield's group"5 who had developed an

expert system for diagnosis in breasthistopathology.The relation between cytological features

and the diagnosis is clearly associated withuncertainty. In Bayesian models this uncer-

tainty is represented by conditional probabili-ties which express the likelihood of a featurebeing present given the diagnosis. In theory, itmight be possible to use statistically definedprobabilities-from frequency counts-toconstruct the CP matrices necessary forBayesian belief networks. In the current

study, however, CP matrices were con-structed using probabilities proposed by thecytopathologist-that is, "personal probabili-ties". They are an attempt to convey in quan-titative terms the knowledge and experienceof the cytopathologist and as such, provideuseful information to a knowledge based sys-tem of this kind. CP matrices are central tothe success of a belief network and care mustbe taken in their construction. If the definedconditional probabilities do not provide theexpected results they can be adjusted in atraining process, based on specimens ofknown diagnosis, until the network performscorrectly. In this particular study, however,the original CP matrices devised by thecytopathologist remained unchanged evenafter testing.To start the network, prior probabilities

must be given for the diagnostic outcomes,benign and malignant. In practice, the atti-tude adopted by most cytopathologists is thateach aspirate is examined equally and inde-pendently, with no preconceived notions as toits most probable diagnosis, regardless of thefact that in statistical terms an individual case ismore likely to be benign than malignant. Wetherefore used equal prior probabilities(05,0O5) for the diagnostic outcomes ofbenign and malignant in the current networkand this was shown to work well.

It was evident that while the results of thefinal belief scores were useful, additionalinformation could be obtained by plotting thecumulative diagnostic probability curve aftereach evidence node had been sampled. Thisallowed the cytopathologist to trace the stepsmade in the diagnostic decision sequence andtheir effect on the final decision, a useful pro-cedure, particularly for difficult cases. It alsohighlighted evidence which did not match thegeneral trend of the curve. If doubt existedover the response of the curve to the evidenceentered for a particular feature then this fea-ture could be reviewed cytologically. Thisquantitative feedback on the diagnosticprocess is fundamental to ensuring that eachpiece of diagnostic evidence is considered infull and that diagnostic accuracy is improved.The storage of diagnostic probability curves inthis way provides a quantitative record of thediagnostic decision process and would haveimportant implications in assessing repro-ducibility and quality control. It is importantto remember that the quantitative characteris-tics of the probability curve are stronglydependent on the order in which the evidencenodes are sampled. This needs to be stan-dardised, as in this study, if comparisons are

Figure 6 Cumulativeprobability graph formalignant lobularcarcinoma.

0E0

0

-W

0)C0

.0

.0

1-009-0-80-70-60-510-40-30-20-1A L

334



j.com/

J Clin P


ownloaded from

http://jcp.bmj.com/


to be made. It should be emphasised, how-ever, that the final probability is completelyindependent of the order in which the evi-dence nodes are examined.The ability of the network to allocate cor-

rectly benign and malignant breast aspirateson the basis of the diagnostic probability wasexcellent and when apparent misclassifica-tions occurred, it was due to lesions which areknown to pose diagnostic difficulties. Onlyone case of fibroadenoma (out of four) fellwithin the equivocal belief range. Supportiveevidence of the benign nature of this case,however, came from the probability curvecharacteristics. The fact that bare nuclei aresuch an important discriminant in the differ-ential diagnosis of fibroadenoma and carci-noma perhaps justifies assigning a strongerCP matrix for this feature in future. The diag-nostic clue binary distance measurement alsoshowed this case to be more closely akin tothe other benign cases than malignancy.The single malignant case which was classi-

fied into the equivocal range by the networkwas an aspirate from an infiltrating lobularcarcinoma. Examination of the cumulativeprobability curve for this case (fig 7) alsoclearly showed the presence of conflicting evi-dence. It is encouraging that the network wascapable of firstly allocating this case in theequivocal (M) belief range and secondly, thehigh C-score would alert the user to theunusual nature of the evidence.

Atypical probably benign cases all fell intothe benign belief range and could not be dif-ferentiated from true benign cases. In practicethis could be satisfactory as atypia is oftenseen in examples of benign epithelial hyper-plasia. All of the atypical cases were benign onbiopsy.

Suspicious cases could be characterised notonly by their equivocal probabilities but bytheir probability curves which were highlyerratic, resulting in high C, T, and P values.Success of the network diagnostic probabilitywas also evident as suspicious cases whichwere classified as EQUIVOCAL (M) orMALIGNANT by the network were con-firmed to be malignant on biopsy.The reproducibility of individual likelihood

ratios for each feature was questionable, butthe final classification of cases in the repeatedstudy was highly reproducible. This demon-strates the need for multiple evidence inreaching a diagnostic decision and shows thatno single feature provides decisive evidence.Bayesian belief networks permit the combina-tion of features which independently may beweak but cumulatively lead to a decision withhigh certainty. While people are capable ofcombining evidence to support a decision,this is often undermined by conceptual, ter-minological, and behavioural inconsistenciesand they cannot quantify the process leadingto, and the certainty of, the final decision.Bayesian belief networks, while maintainingthe role for linguistic and descriptive termi-nology for evidence communication, providea quantitative means of describing the diag-nostic decision making process.

These results verify the potential of this sys-tem as an aid to decision making. The net-work needs to be tested on a larger series ofcases with emphasis on those cases whichpose particular difficulties-for example, lob-ular carcinomas. Only by examining largernumbers of difficult cases can methods toimprove network performance be found.

In the current design only cytological evi-dence has been used. In practice, clinical evi-dence provides important supplementary dataused in the diagnosis. Future work will incor-porate clinical evidence either within theexisting network structure or as a separatenetwork whose result can be used to supportor reject the outcome of the cytological beliefnetwork. Also the network described here isbased on features seen in air-dried slides but asimilar network could be constructed usingfeatures appropriate for alcohol fixed prepara-tions.The value of the system designed in the

current study is three-fold. Firstly, it can beused to provide the cytopathologist with anaid to decision making in the routine diagnos-tic assessment of breast aspirates. At themoment the design of the system is primitivein that the cytopathologist has to providenumbers (likelihood ratios) for each evidencenode. This can be time consuming and evenwith practice can still take between five and10 minutes per case. This could be overcomeby using stored image libraries with pre-defined likelihood ratios, allowing the user toselect images which most closely match thefeatures seen in the current case. Thisdivorces the user from having to convert his orher cytological impressions into numbers yetprovides a highly visual means of providingdata suitable for the network. As cumulativeprobability curves for cases with certain out-comes are recorded, statistical data willbecome available on conditional probabilityvalues between P, T, and C scores and thediagnostic outcome. The network could beextended to use this information about itsown decision making process as additionaldiagnostic clues, and enter likelihood vectorsas a last step before it reports a final outcomeprobability.

While an interactive system as describedmight prove successful, the role of quantita-tive cytomorphometric data should not benegated. Numerous studies have shown thatnuclear and cellular measurement can be ben-eficial in the objective diagnosis of breastcytologyl'20 and future work should attemptto include these as additional evidence withinthe framework of a Bayesian belief network.

Secondly, a model of this sort can be usedas the basis for a teaching system for thoseinexperienced in breast FNAC diagnosis. Thecomputer could, therefore, act as a consul-tant, leading the user through the diagnosticprocess, prompting for information about thecase in hand, providing visual examples ofdiagnostic features, querying the users'responses and on reaching a decision, retracethe steps used to obtain it.

Finally, it would be desirable to design a

335



j.com/

J Clin P


ownloaded from

http://jcp.bmj.com/


computer system capable of automated diag-nosis in breast FNAC. Such systems havebeen considered for many years in cervicalcytology2l-23 and recent increases in processingpower and advances in machine vision, expertsystem, and neural network technology havenow made this a reality.

It is clear that expert computer systems are

going to be increasingly investigated in diag-nostic histology and cytology in the next fewyears. Computerised interactive decision sup-port systems, such as the one described here,will, without doubt, lead to improved consis-tency and reproducibility in diagnosis.

This study was supported by the Ulster Cancer Foundation(PWH), The Queen's University of Belfast, and by a grantfrom the National Institutes of Health, Bethesda, Maryland,USA, R 35 CA 538701 (PHB).Dr Neil Anderson held an RVH fellowship during the develop-ment of this project.

1 Smeulders AWM, van Ginneken AM. An analysis ofpathology knowledge and decision making for the devel-opment of artificial intelligence-based consulting sys-tems. Anal Quant Cytol Histol 1989;11:154-65.

2 Bartels PH, Hiessl H. Expert systems in histopathology. IIKnowledge representation and rule based systems. AnalQuant CytolHistol 1989;11:147-53.

3 van Ginneken AM, Smeulders AWM. Reasoning in uncer-tainties: An analysis of five strategies and their suitabilityin pathology. Anal Quant Cytol Histol 1991;13:93-109.

4 Bartels PH, Thompson D, Weber JE. Expert systems inhistopathology. IV The management of uncertainty.Anal Quant Cytol Histol 1992;14:1-13.

5 Bibbo M, Bartels PH, Pfeifer T, Thompson D, Minimo C,Galera Davidson H. Belief network for grading prostatelesions. Anal Quant Cytol Histol 1993;15:124-35.

6 Montironi R, Bartels PH, Thompson D, Scarpelli M,Hamilton PW. Prostate intraepithelial neoplasia: devel-opment of Bayesian belief network for diagnosis andgrading. Anal Quant Cytol Histol (in press).

7 Bartels PH, Thompson D, Bibbo M, Weber JE. Bayesian

belief networks in quantitative histopathology. AnalQuant Cytol Histol 1992;14:459-73.

8 Morawski P. Understanding Bayesian belief networks. AlExpert 1989;August:44-8.

9 Morawski P. Programming Bayesian belief networks. AlExpert 1989;May:74-9.

10 Pearl J. Probabilistic reasoning in intelligent systems.California, Morgan Kaufman, 1988.

11 DoH. NHS Breast Screening Programme. Guidelines forCytology Procedures and Reporting in Breast CancerScreening. NHSBSP Publication No 22. London: HMSO,1992.

12 MacWilliams Hi, Sloane NJA. The theory of error-correctingcodes. Part I. Amsterdam: North Holland PublishingCompany, 1977.

13 Heathfield HA, Kirkham N, Ellis IO, Winstanley G.Computer assisted diagnosis of fine needle aspirate ofthe breast. Y Clin Pathol 1990;43:168-70.

14 Bartels PH, Thompson D, Weber JE. Diagnostic decisionsupport by inference networks. In Vivo 1993;7:379-86.

15 Heathfield H, Bose D, Kirkham N. Knowledge-basedcomputer system to aid in the histopathological diagnosisof breast disease. Y Clin Pathol 1991;44:502-8.

16 Cornelisse CJ, de Loning HR, Arentz PW, Ratgever JW,van Heerde P. Quantitative analysis of the nuclear areavariation in benign and malignant breast cytology. AnalQuant Cytol 1981;3:128-34.

17 Boon ME, Trott PA, van Kamm H, Kurver PJH, Leach A,Baak JPA. Morphometry and cytodiagnosis of breastlesions. Virchows Arch (PatholAnat) 1982;396:9-18.

18 Wolberg WH, Tanner MA, Loh W-Y, Vanichsetakul N.Statistical approach to fine needle aspiration diagnosis ofbreast masses. Acta Cytol 1987;31:737-41.

19 van Diest PJ, Risse EK, Schipper NW, Baak JP,Mariquand J. Comparison of light microscopic gradingand morphometric features in cytological breast cancerspecimens. Path Res Pract 1989;185:612-16.

20 Beerman H, Veldhuizen RW, Blok RA, Hermans J, OomsEC. Cytomorphometry as quality control for fine needleaspiration. A study in 321 breast lesions. Anal QuantCytolHistol 1991;13:143-8.

21 Tang Z, Savino A, Wong EK, Koss LG, Shaw LG. Anexpert system designed as a tutoring tool in cervicalcytology: TTCC-1 system. Anal Quant Cytol Histol1988;10:417-22.

22 Hall T. Detection of abnormal cells in cervical smearsusing a neural network-based pattern recognition sys-tem. Anal Quant Cytol Histol 1990;12:194-5.

23 Anonymous. Data on automated cytology systems as sub-mitted by their developers. [Editorial]. Anal Quant CytolHistol 1991;13:300-6.

336



j.com/

J Clin P


ownloaded from

http://jcp.bmj.com/

j expert systemsupport bayesianbeliefgist entered evidence into the networkin the formoflikelihood...

Documents