Combining Artificial Neural Networks and Transrectal Ultrasound in the Diagnosis of Prostate Cancer

October 1, 2003

Prostate cancer management issurrounded by controversy.From the screening debatethrough choosing the best treatmentoption for localized disease, there islittle consensus on the approach to themost common solid tumor in men. Avariety of predictive models are beingdeveloped to assist in clinical decisionmaking.[1,2] Although transrectal ultrasound(TRUS)-directed prostatebiopsies represent the “gold standard”in the diagnosis of the disease, limitationsof this approach have been recognized.[3] To compensate for theselimitations, the absolute number of needlecores taken has increased from 6 to10–12 or more. TRUS enhancementssuch as color Doppler and the use ofcontrast agents hold promise, but theyhave not yet replaced the TRUS grayscaleapproach.[4]

Prostate cancer management issurrounded by controversy.From the screening debatethrough choosing the best treatmentoption for localized disease, there islittle consensus on the approach to themost common solid tumor in men. Avariety of predictive models are beingdeveloped to assist in clinical decisionmaking.[1,2] Although transrectal ultrasound(TRUS)-directed prostatebiopsies represent the "gold standard"in the diagnosis of the disease, limitationsof this approach have been recognized.[3] To compensate for theselimitations, the absolute number of needlecores taken has increased from 6 to10-12 or more. TRUS enhancementssuch as color Doppler and the use ofcontrast agents hold promise, but theyhave not yet replaced the TRUS grayscaleapproach.[4]Drs. Porter and Crawford have reviewedone strategy that may improvethe diagnostic accuracy of prostatecancer biopsy and potentially avoidunnecessary biopsies-artificial neuralnetworks (ANNs) combined withTRUS biopsy as a predictive model.Artifical Neural Networks
The ANN is a specific type of computer-driven, artificial intelligencesoftware system, based on the neuralstructure and function of the brain.Basic processing units called nodessimulate neurons, and weighted interconnectionsbetween the nodes simulatethe basic units of the nervoussystem, dendrites, and axons.[5] Theseinterconnections weigh functions asmultipliers that simulate the connectionstrengths in the analogousbiologic model. ANNs are not programmed,but are unique in that most"learn" by experience, ie, a "supervisedlearning" phase known as training.The ANN learning set consists ofactual clinical case inputs and knownoutputs (ie, results). These clinicalvariables (in this case, data such asdigital rectal exam findings, prostatespecificantigen [PSA] levels, and age)and a known pathologic outcome arepresented to the ANN sequentially andrepeatedly. A training algorithm automaticallyadjusts the connectionweights, consequently changing theoutput values, to reduce the errorsbetween the actual ANN outputs andthe expected outputs. With training, aset of connections is developed to allowfor the greatest number of correctpredictions for a given training dataset.Measuring ANN Performance
Next, the system is validated withnew cases to determine the accuracyof the prediction. The performance ofan ANN can be measured by calculatingthe sensitivity, specificity, andboth negative- and positive-predictivevalue of a specific ANN output. Theoverall performance may be quantifiedby generating a receiver-operatorcharacteristic (ROC) curve.TRUS Images Validate Output
ANNs have great potential in theoverall management of prostate cancer.[6,7] Though not directly addressedin this review, another use for ANNtechnology is to enhance TRUS imagesand identify malignant foci. In fact,one of the earliest reports of the use ofANNs in prostate cancer, published in1992, involved the analysis of TRUSimages of the prostate.[8] In this preliminarywork, the ANN was able todifferentiate between prostatic and nonprostatictissues in TRUS images. Pureultrasound radiofrequency images inputtedinto the ANN model for theanalysis of ultrasound images of thebreast, colon, prostate, and other tissueshave shown strong results in theidentification of malignancy.[9-11]By gathering actual TRUS imagesprior to radical prostatectomy andcomparing these to whole-mount pathologyslides, prospective librariesof prostate tissue image types havebeen developed and incorporated intoANNs. With such a technique, anANN was used to identify areas suspiciousfor cancer in a validation setof TRUS images.[10] Preliminary datademonstrated that 99% of confirmedbenign samples and 79% of malignantlesions were correctly classified.Of isoechoic cancers, 97% were correctlyclassified by the ANN.This technology is also applicableto other modalities such as magneticresonance and color Doppler imaging.[4,12] Automated image analysis,including pattern recognition and ANNsapplied in real time to TRUS images,may successfully identify lesions thatcannot be seen by the human eye.Such automated image analysis andpattern recognition, however, is currentlyunavailable for TRUS systems.Predictive Modeling Tools
As noted, predictive modeling toolsare being developed to assist the clinicianin the decision-making process.Traditionally, these tools have reliedon solely clinical parameters (eg, PSA,clinical stage). ANN technology allowsthe integration of many morecomplex variables into this decisionprocess. Although specific identificationof malignancies on imagingstudies was the first use of ANNtechnology, Drs. Porter and Crawfordreview the combination of ANNs withbasic clinical parameters and ultimatediagnostic TRUS imaging and biopsyto develop predictive models for prostatecancer needle biopsy. This couldbe the next major advance in our abilityto avoid unnecessary prostate biopsies.The integration of clinicalparameters or other markers to improvethe identification of prostate cancer is abasic hurdle that must be overcome.While these developments reviewedby Porter and Crawford areoutstanding, they are not the final solutionto the prostate cancer screeningand diagnosis controversy. Manymen have clinically insignificant prostatecancer at the time of diagnosis. Atheoretical concern is that this elegantANN technology may inadvertentlylead to overdiagnosis of so-called autopsycancers. Indeed, the diagnosisof the "clinically insignificant," oftensmall-volume prostate cancers is oneof the major arguments against screeningfor prostate cancer. These insignificantprostate cancers may nevercause clinical symptoms nor result indeath ("more men live with than diefrom prostate cancer") and might beamenable to watchful waiting approaches.[13] Whether the enhancedmethods proposed by the authorscause more insignificant tumors to bedetected remains to be seen.In addition, we need a technologythat can differentiate the life-threateningtumors from the indolent oneseven before the prostate is violated bya needle through the rectum. CouldANNs be incorporated into an imagingmodality that would allow theidentification of tumors and their classificationas insignificant or significant?Although preliminary work hasbeen done in identifying the aggressivenessof prostate cancer on imagingstudies, this remains a challengefor the future.[14]Predictive models are becomingcommonplace in the management ofprostate cancer. The benefit of theANN described by the authors is thatmore complex variables can be analyzedto help with decision-making inthis complex disease.

Disclosures:

The author(s) have no significant financial interest or other relationship with the manufacturers of any products or providers of any service mentioned in this article.

References:

1.

Ross PL, Scardino PT, Kattan MW: Acatalog of prostate cancer nomograms. J Urol165:1562-1568, 2001.

2.

Eastham JA, Kattan MW, Scardino PT:Nomograms as predictive models. Semin UrolOncol 20:108-115, 2002.

3.

Presti JC Jr: Prostate biopsy: How manycores are enough? Urol Oncol 21:135-140, 2003.

4.

Ismail M, Gomella LG: Ultrasound forprostate imaging and biopsy. Curr Opin Urol11:471-477, 2001.

5.

Dayhoff J: Neural Network Architectures:An Introduction, pp 70-74. Boston, InternationalThomson Computer Press, 1996.

6.

Errejon A, Crawford ED, Dayhoff J, et al:Use of artificial neural networks in prostatecancer. Mol Urol 5:153-158, 2001.

7.

Ziada AM, Lisle TC, Snow PB, et al:Impact of different variables on the outcome ofpatients with clinically confined prostate carcinoma:Prediction of pathologic stage and biochemicalfailure using an artificial neuralnetwork. Cancer 91(8 suppl):1653-1660, 2001.

8.

Prater JS, Richard WD: Segmenting ultrasoundimages of the prostate using neural networks.Ultrason Imaging 14:159-185, 1992.

9.

Noritomi T, Machi J, Feleppa EJ, et al: Invitro investigation of lymph node metastasis ofcolorectal cancer using ultrasonic spectral parameters.Ultrasound Med Biol 24:235-243,1998.

10.

Loch T, Leuschner I, Genberg C, et al:Artificial neural network analysis (ANNA) ofprostatic transrectal ultrasound. Prostate 39:198-204, 1999.

11.

Lizzi FL, Astor M, Feleppa EJ, et al:Statistical framework for ultrasonic spectral parameterimaging. Ultrasound Med Biol 23:1371-1382. 1997.

12.

Miller SM, Ackermann R: Color Dopplersonography of the prostate. Urol Int 57:158-164, 1996.

13.

Klotz L: Expectant management withselective delayed intervention for favorable riskprostate cancer. Urol Oncol 7:175-179, 2002.

14.

Ismail M, Petersen RO, Alexander AA,et al: Color Doppler imaging in predicting thebiologic behavior of prostate cancer: Correlationwith disease-free survival. Urology 50:906-912, 1997.