TheArtificialNeuralNetworksforCancerResearchin

Prediction&Survival(ANNCRIPS)
KaranKamdar*,AmitMathapati**
* VivekanandInstituteofTechnology,Mumbai,India
** VivekanandInstituteofTechnology,Mumbai,India

karan@stople.com amitmathapati@yahoo.com

Abstract Prostate cancer is the most common cancer

among men in age 50 or older, excluding skin cancer.
Current methods of screening for prostate cancer carried
out through blood tests & presence of high PSA lead to a
high percentage of false positive test result (FPTRs) which
can be reduced by employing intelligent Artificial Neural
Networks. The goal of our research paperand the parallel
undertaking of its practical implementation is to develop a
mathematical model to improve prostate cancer detection
and staging systems and finally to present a deploy ready
marketable solution based on the model which can be
installed across various screening centers, hospitals and
researchorganizations.

I.

tobeaccompaniedwithadigitalrectalexamwherea
biopsyoftheprostatemayberequiredorsomeother
methodtoconfirmthepresenceofcancer.Thesemethods
workinsituationswherethereisstrongverifiable
evidenceof thepresenceofcancer.Howeverinextreme
caseswherethereisnoclearsignalofcanceroccurrence
ahighnumberoffalsepositivetestresults(FPTRs)may
beproducedresultinginrepeatedscreeningand
diagnosticprocedures.Hence,thegoalofthispaperisto
employIntelligentArtificialNeuralNetworksto builda
deployready marketablemathematicalmodeltoreduce
thenonrequiredtrialanderrormethodsandexpedite
earlycancerdiagnosisandtreatment

INTRODUCTION
II.

A. TheGenesisWhyistheemphasisonprostate
cancer?
Prostatecanceristhemostcommoncanceramongmen,
excludingskincancer.
AmericanCancerSociety(ACS)estimatesfor2005
include232,090newcasesofprostatecancerintheU.S.
Year2005estimatesinclude30,350deathsoccurring
fromprostatecancerintheUSalone,makingitthe
secondleadingcauseofcancerdeathinmen.
Allmenareatriskforprostatecancer.Theriskincreases
withage,andfamilyhistoryalsoincreasestherisk.
AfricanAmericanmenhaveabouta70percenthigher
incidencerateofprostatecancerthanCaucasianmen,and
nearlyatwofoldhighermortalityratethanCaucasian
men.
B. TheNeedforefficientprostatecancerscreeningand
diagnosticmethods
Increasingly,physiciansarerelyingontheresultsfrom
thePSA(prostatespecificantigen)bloodtesttodiagnose
prostatecancer.Whentheprostateglandhasahintof
cancerproducingcells,adetectablesubstance,namely
PSA,isproducedintheblood.AnabnormallyhighPSA
levelalertsthephysiciantothepossiblepresenceof
prostatecancer.AlthoughthePSAbloodtesthasbeen
refinedanddeveloped,thephysiciancannottrulyrelyon
theresultsofthistestaloneandsothePSAbloodtesthas

THE ANNCRIPS

A. ProjectOverview
TheArtificialNeuralNetworksforCancerResearchin
PredictionandSurvival(ANNCRIPS)isavoluntaryeffort
startedbytheauthorsofthispaper.Theideaistobuilda
mathematicalmodelto improveprostatecancerdetection
andstagingsystems.Herethebasis istoaddintelligence
toArtificialNeuralNetworksandproduceadeployready
mathematical model that revolves around the concept of
ANNs. The actual implementation of the model would
require building a standalone software application which
would be installed across various screening centers,
hospitals and research institutions. The project is in its
intermediate stage and the team is working on the
development and implementation stages to bring the
product to the masses. C is the choice of programming
language to be used for the reasons of its strong
mathematicallibraryandMatlabisusedasthesimulation
software packagetotestoursoftware runs.This research
project is a long term initiative involving successive
refinementstothefirstversionofourmodel.
B. ArtificialNeuralNetworks

D. LinkingANNstoProstateCancerAnalysis
TheArtificialNeuralNetworks(ANNs)areinformation
processing paradigms inspired by the way biological
nervous systems such as the brain process information.
Thefigure1.1showssuchatypicalexampleofthehuman
brains neural network. ANNs adopt this interconnected
neuron network to perform complex computations. The
typicalANNmodelisasshown below:

Fig 1.2

The current method for prostate cancer detection is

finding the percentage of tPSA (total prostate specific
antigen).Howevertherearesomecoredrawbacksrelated
tothisapproach.Thefirstbeingthatonlycancerconfined
totheprostateglandcanbedetectedandthesecondbeing
that additional information such as the degree of cancer
spread to other internal organs requires more tissue
samplings.
Apossiblesolutiontotheseproblemsmaybefoundin
finding the percentage of fPSA (free prostate specific
antigen)tosuggestthedegreeofspreadofprostatecancer
tootherparts.Howeverboththeabovesolutionshavethe
followingdrawbacks:
a)They cannotprovide specificdiagnosticresultsona
persontopersonbasis
b) There are unable to learn from existing diagnostic
patterns and associate them with new input data from
differentpatientsfallingunderthesamecategorybasedon
age group, degree of cancer spread, number of years till
cancerwasdiagnosedandothergenericparameters.
c) Both the systems provide accuracy levels ranging
from satisfactory to moderately efficient and hence
repeateddiagnosticmeasuresto overcometheoccurrence
offalsepositivetestresultshavetobeundertaken.
E. BuildingtheANNmodel

C. OverviewofProstateCancer

OurANNModelisbasedontheconceptonMultilayer
Perceptron(MLP).Itconsistsofanetworkofprocessing
elementsornodesarrangedinlayers.
Principle Input pattern presented at the input layer
causes network nodes to perform calculations in the
successive layers until an output value is computed at
eachoftheoutputnodesfromwhichthemostsignificant
isselected

Fig 1.3

Theprostateisasexglandinmen.Itisaboutthesizeofa
walnut,andsurroundstheneck ofthebladderandurethra
thetubethatcarriesurinefromthebladder.Itispartly
muscularandpartlyglandular,withductsopeninginto
theprostaticportionoftheurethra. Theprostategland
secretesaslightlyalkalinefluidthatformspartofthe
seminalfluid,afluidthatcarriessperm.
Prostatecanceristhecancerbeginningintheprostate
whichmayremainintheprostateglandorspreadtoother
internalorgans.Screening&Diagnosisarecarriedout
throughbloodtestsmeasuringthelevelofPSAand
alternativelyaccompaniedbyaDREordigitalrectum
exam.Afterinitialdiagnosisfurtherstagingsystemsuch
asTNMmaybeundertakentopinpointthelocationof
cancerandthedegreeofitsspreadintootherregions.

Fig 1.4

Working

Inputto nodej :
Outputofnodej:

This continues through all the layers of the

networkuntiloutputlayerisreachedandoutput
vector is computed The function f denotes

activation function of each node. A sigmoid

functionisgenerallyused

F. ExtendingtheMLPmethodology
Ourneuralnetworkmodelusedisamultilayerperceptron
(MLP) network composed of one input layer with four
primary preprocessed variables (tPSA, fPSA, prostate
volume,DRE).

Afterthisthemodelisrunwithinternalprocessinginthe
networktakingplaceandresultantoutputtakingthevalue
between0(lowPCarisk)and1(highPCarisk).Insome
cases, the value is <0 or >1 which is not relevant.
Multiple runs of the model are now executed and a
statistical curve indicating the values of PCa risk is
plotted. If the curve indicates high occurrence of values
closerto1thenexistenceofprostatecancerisconfirmed.
Theresultantdataisnotexcretedbutisfeedbackintothe
system so that a higher accuracy of the statistical curve
applicable to patients belonging to similar data set is
madepossible.
Intelligentlearningandnetworktrainingisthekeytoour
ANNmodelssuccess.Ifthisisnotdonethenundesirable
results may be obtained. E.g. if the network has been
trainedonlyonceforapersonhavinghighPSAleveldue
to noncancerous cells then it wont indicate any cancer
riskforpersonsfallingintothesamecharacteristicgroup
and actually having a high PSA level. Hence repeated
trainingandlearningshould be employedwidelysothat
themodellearnstoassociateanddeassociateitselfatthe
sametimefromdifferentpatternsets.Themodelcanalso
be categorized as stochastic for the reasons of having
input data set which is variable across different patients
whoaretested.

Fig 1.5

There is one hidden layer with two neurons, and one

outputlayerwithoneneurongivingtheoutputvaluethat
isameasureoftheprobabilityofcancer.
The 13 parameters (weights) associated with each input
layer are to be optimized and the Activation function
usedinthehiddenandoutputlayersishyperbolictangent
sigmoidfunction.

G. TrainingProcedure
The following training procedure is used for our MLP
basedANNmodel

[a=tanh(s)=(e^se^s)/(e^s+e^s)]
Hencetheoutputvaluesarebetween1to1
Theformulaforwholenetwork thusbecomes:
a1=tanh(IW1,1*X1+IW2,1*X2+IW3,1*X3+
IW4,1*X4+IB1)
a2=tanh(IW1,2*X1+IW2,2*X2+IW3,2*X3+
IW4,2*X4+IB2)
aout=tanh(LW1*a1+LW2*a2+LB)
After model buildup, the respective weights are
initialized and the inputs from the patient data set are
entered. Inputs are actual dataofpatients suffering from
prostatecancer. The whole networkisnowtrainedusing
LevenbergMarquardt & Bayesian optimization
techniques to recognize and associate patterns of input
withdesiredoutputsindicatingthecorrectcancerrisk.

Fig 1.6

H. 5StepstoimplementtheANNCRIPSmodel
n Step1 ObtainingtheInputDataSet

From the above graphical chart it is seen clearly that

our ANN model would better the fPSA or tPSA models
quitesubstantially.

ROC Curve : Our ANN model also occupies a

greaterareaundercurveascomparedtofPSA&
tPSAmodels

Fig 1.7

n Step2 AdjustingInitialWeights
i) The model is fitted on 75% and tested on 25%
patientsin eachtrainingset.
ii) Randomization, fitting and testing sessions
repeated 56 times andthe weights producing smallest
sumofsquarederrorsontheinitialtestsetareselectedas
initialweightsforthefinaltrainingoftheMLP.
n Step3 TraintheNetwork
ActualTrainingofNetworkbeginswithestimation&
validationsamples.
n Step4 TesttheNetwork
If validation objective is acquired, network trained
againtillbestperformanceparametervaluesareobtained.
n Step5 Validating&Generalizing
NetworkEfficiency

Fig 1.8

III. THEPROPOSEDANNCRIPS PRODUCT

Thefollowingisasampleofhowourproductwouldlook
likeonceithasbeenimplemented.Theteamiscurrently
workingonproducingasoftwaresolutionbymappingthe
proposed mathematical model to the C language and
building simulation runs through Matlab to exploit the
core mathematicalfunctionality of the two programming
packages. We also find tremendous scope in terms of
improvement in the mathematical model, training
procedure,testingandvalidationsteps.

I. ResultsofourANNonacomparativebasis

Fig 1.9

Sensitivity :Abilityofthemodeltodetect
prostatecancerearly.
Specificity :Efficiencyofavoidingrepeated
tissuesamplings.
Accuracy :Proportionofsubjectswithacorrect
testresult.

Fig 2.0

IV. CONCLUSION
l Prostate cancer is the most common form of
cancer among men according to statistical study
results.
l Screeningisaveryroughestimateforcancerrisk
l Further staging systems such as A, B, C, D,
TNMlackrequiredefficiency.
l Solution: Our proposed ANN model is far more
efficient in predicting prostate cancer risk and
reducingtheno.offalsepositivetestresults.
l Our model is developed to aid or substitute the
currentdiagnosisandprognosismethods.
l Our software test runs have shown that highly
accurate ANNs based on extending the MLP
model can detectprostatecancerearly&reduce
unnecessary tissue samplings as compared to
currentmethodssuchasfPSAandtPSA.

V.

ACKNOWLEDGEMENT

Theteamwouldliketo thankProf. Balakrishnan who

is the Head of Computer Science Department at the
Vivekanand Institute of Technology, Mumbai for his
guidance and support when we initially looked for
references and sources of information on this research
project. We look forward to his guidance in the years to
come.WewouldalsoliketothanktheStudentLibraryat
the Vivekanand Institute of Technology, Mumbai for
providing exhaustive resources for Prostate Cancer
ResearchandNeuralNetworkreferences.

Influencingfactors:
l Larger no. of input variables from the patient
datasetcanbesupported.
l Interconnectingrelationshipsbetweentheseinput
variables and thereby formation of reusable
patternscanbeestablished.
l The ability of the neural network to be trained
time and again can be exploited thereby
increasing accuracyeachtime.
l A deployready easy to install software solution
would be made available in the form of our
ANNCRIPS product which can installed across
variousscreeningcenters,hospitalsandresearch
institutions.

REFERENCES
[1] ForResearchonNeuralNetworksandProstateCancer:
VivekanandInstituteofTechnologysStudentLibrary,
http://www.vesit.edu
[2] ForStatisticsonprostatecancer:
http://www.health.uab.edu/show.asp?durki=15376.