15
1 PROSTATE CANCER: PROSTATE CANCER: The Diagnostic Dilemma The Diagnostic Dilemma Christopher R. Porter, MD Co-Director Urologic Oncology Virginia Mason Medical Center Seattle, WA The Dilemma Common disease Sensitive, non-specific screening test Almost everyone gets a biopsy Tumor is invisible It’s a shot in the dark… and hopefully not a miss! In an Ideal World Who to shoot and where to shoot them We would know… Who needs a biopsy and where to aim OR

PROSTATE CANCER: The Diagnostic · PDF file1 PROSTATE CANCER: The Diagnostic Dilemma ... zInputs: PSA/VOL/DRE/TRUS/BX HX ... Djavan B 272 ANN/LR Eastham JA 700 LR Age,

Embed Size (px)

Citation preview

Page 1: PROSTATE CANCER: The Diagnostic  · PDF file1 PROSTATE CANCER: The Diagnostic Dilemma ... zInputs: PSA/VOL/DRE/TRUS/BX HX ... Djavan B 272 ANN/LR Eastham JA 700 LR Age,

1

PROSTATE CANCER: PROSTATE CANCER: The Diagnostic DilemmaThe Diagnostic Dilemma

Christopher R. Porter, MDCo-Director Urologic OncologyVirginia Mason Medical Center

Seattle, WA

The Dilemma

Common diseaseSensitive, non-specific screening testAlmost everyone gets a biopsyTumor is invisible

It’s a shot in the dark…

and hopefully not a miss!

In an Ideal World

Who to shoot and where to shoot them

We would know…

Who needs a biopsy

and where to aim

OR

Page 2: PROSTATE CANCER: The Diagnostic  · PDF file1 PROSTATE CANCER: The Diagnostic Dilemma ... zInputs: PSA/VOL/DRE/TRUS/BX HX ... Djavan B 272 ANN/LR Eastham JA 700 LR Age,

2

A Big Deal

Why is prostate cancer so important?

Because a lot of people have it…

Worldwide Incidence

*Incidence per 100,000 population. Parkin DM, et al. CA Cancer J Clin. 1999;49:53.

39.5539.55

16.7516.75

8.518.51

49.7049.70

1.081.08

5.135.13

31.0331.03

92.3992.39

Eastern Eastern EuropeEurope

JapanJapan

AustraliaAustraliaNew ZealandNew Zealand

ChinaChina

Northern Northern AfricaAfrica

Southern Southern AfricaAfrica

North North AmericaAmerica

WesternWesternEurope Europe

Incidence in US

230,000 new cases in 2004*28,900 deaths in 2003#1 cancer and #2 killer in men4.5% decline per year since 1994

*Jemal A, Tiwari RC, et al CA Cancer J Clin 2004;54:8-29.

Screening Tools

Oesterling J, et al. Cancer: Principles & Practice of Oncology. 5th ed. 1997;1322-1386.

MethodMethodSensitivitySensitivity

(%)(%)SpecificitySpecificity

(%)(%)

PositivePositivePredictivePredictiveValue (%)Value (%)

DRE 69-89 84-98 26-35

PSA* 57-79 59-68 40-49

TRUS 36-85 41-79 27-36

*Using a 4-ng/mL cutoff.

Prostate Specific Antigen (PSA)

<4 ng/mL considered “normal”4-10 ng/mL 22% positive biopsy rate>10 ng/mL 66% positive biopsy rateMay be elevated by any prostate disease or manipulation

Oesterling J, et al. Cancer: Principles & Practice of Oncology. 5th ed. 1997;1322-1386. Kassabian VS, et al. The American Cancer Society Textbook of Clinical Oncology. 2nd ed. 1995;311-329.

Brawer MK. CA Cancer J Clin. 1999;49:264-281.

Page 3: PROSTATE CANCER: The Diagnostic  · PDF file1 PROSTATE CANCER: The Diagnostic Dilemma ... zInputs: PSA/VOL/DRE/TRUS/BX HX ... Djavan B 272 ANN/LR Eastham JA 700 LR Age,

3

ROC Curve for PSA

Bellorofonte C, et al: European Urology 2005;47:29-27.

The problem is that PSA is sensitive but not specific

PSA 4 -10 ng/dlPositive biopsy: 24%

CaP Prevention Trial: Results

FINASTERIDE24.8% decrease in CaP

“Higher grade” CaP (6.4% vs 5.1%)

Sexual side effects slightly more common

(BUT high in placebo too! (50%))

24.4% with PSA <3 had positive biopsy

Surprisingly…

One in four men with PSA <3 have CaP

The prostate cancer prevention trial: current status.Higgens et al J Urol 2004;171(Part 2):1517.

Consider the Cost…

Page 4: PROSTATE CANCER: The Diagnostic  · PDF file1 PROSTATE CANCER: The Diagnostic Dilemma ... zInputs: PSA/VOL/DRE/TRUS/BX HX ... Djavan B 272 ANN/LR Eastham JA 700 LR Age,

4

Consider the Cost

8% have PSA >4¼ will be positive US: 38 million men age 45-748% of 38 million = 3.04 million¼ of 3.04 million = 760,000 prostate cancer2.28 million unnecessary biopsies

Grizzle WE et al, J U Onc. 2004; 22: 337-43

It gets worse…

Consider the Cost

The Prostate Cancer Prevention Trial (PCPT)Placebo Arm

After 7 years those with PSA <4 (-DRE)15% positive end of study biopsy

and Re-consider the Cost

If we screen EVERYONE:38 million only 8% (3 million) screened35 million remain15% of these will be positive (5.24 million)Total number with cancer 6 million

32 million unnecessary biopsies!!

What can we do to PSA to improve accuracy of cancer

detection?

Page 5: PROSTATE CANCER: The Diagnostic  · PDF file1 PROSTATE CANCER: The Diagnostic Dilemma ... zInputs: PSA/VOL/DRE/TRUS/BX HX ... Djavan B 272 ANN/LR Eastham JA 700 LR Age,

5

Improving PSA

Lower the cut-off pointUse Free/Total PSA ratioUse complex PSAUse precursor forms of PSA

Free-to-Total PSA*

PSAPSA Probability of CancerProbability of Cancer

2 ng/mL 1%

2-4 ng/mL 15%

4-10 ng/mL 25%

>10 ng/mL >50%

Brawer MK. Prostate-specific antigen: Current status. CA Cancer J Clin. 1999;49(5):264-281.

% FPSA% FPSA Probability of CancerProbability of Cancer

0-10% 56%

10-15% 28%

15-20% 20%

20-25% 16%

>25% 8%*Men with non-suspicious DRE results, regardless of patient age.

Free PSA ROC Curve

Bellorofonte C, et al European Urology 2005;47:29-27.

Even with the most advanced use of molecular forms, one still cannot predict who will have a positive biopsy.

Using PSA Forms to Improve Accuracy

Free PSA at least 3 precursor forms:• A full length & two truncated• Precursors can be expressed as % of free PSA

(%pPSA)• Accuracy of %pPSA vs % free PSA is 64% vs 53%

Improving PSA

PSA densityPSA velocity Free PSAPSA precursor forms etc

The fact of the matter is…

They are really not much better.

Page 6: PROSTATE CANCER: The Diagnostic  · PDF file1 PROSTATE CANCER: The Diagnostic Dilemma ... zInputs: PSA/VOL/DRE/TRUS/BX HX ... Djavan B 272 ANN/LR Eastham JA 700 LR Age,

6

The Problem

INDIVIDUAL CLINICAL

PARAMETERS ARE UNRELIABLE

Gestalt

MENTAL PREDICTION

Personal experience

Personal knowledge

Prior outcomes

Mental prediction/human decision-making subject to inherent biases

Hogarth, R: Judgment & Choice, The Psychology of Decision. 2nd Ed. New York. John Wiley & Sons

Predictive Models

Not subject to same inherent biasesMaximize predictive accuracyOutperform human experts

Miehl P, Causes & Effects of my Disturbing Little Book. J Pers. Asses. 50: 370, 1986Larkin, M, Cancer Prognostic tools..Oncol Times, April,2002, p58

Predicting the Outcome of Prostate Biopsy in a Racially Diverse

Population: A Prospective Study

American Society of Clinical Oncology Orlando, Florida, 2001

Porter CR, et al.

Contributors

Colin O’DonnellE David CrawfordEduard GamitoChristopher PorterAshutosh Tewari

DOD Grant: ANN’S in CaP Project

Methods

Prospective study (1999-2001)

IRB approval

319 men: Indication for TRUS Biopsy

Single attending surgeon

Clinical & Ultrasonic Exams

DRE & TRUS: Level of Suspicion, 1-5

Page 7: PROSTATE CANCER: The Diagnostic  · PDF file1 PROSTATE CANCER: The Diagnostic Dilemma ... zInputs: PSA/VOL/DRE/TRUS/BX HX ... Djavan B 272 ANN/LR Eastham JA 700 LR Age,

7

Methods

AgeRaceFamily HistoryAUASSBiopsy HistoryPotency

PSAFree PSAPSA DensityProstate VolumeDRE (LOS)TRUS (LOS)TRUS: Calcium

PARAMETERS:

Methods

STATISTICSUnivariate Analysis

PREDICTIVE MODELLINGLogistic Regression Analysis (LR)Artificial Neural Networks (ANN)

Methods

Pre-biopsy parametersSelection: Stepwise LRInputs: PSA/VOL/DRE/TRUS/BX HXModel training

Five Logistic Regression models *Five ANN models ^

* 1.2.1 R Development Program^ Brainmaker version 3.72 Calif. Sci 1988

ANN Model Training

Data set divided: Training & Validation Set

Cases with known output are presented

Training algorithm adjusts the weights

Weights based on error actual vs. expected

Training set data presentation continues

Weights adjusted to minimize error

ANN Model validated on “virgin data”

YDRE + BX

Neural Networks

PSA

TRUS

XInput Layer Hidden 1 OutputHidden 2

Receiver Operator Characteristics (ROC) Curve

Measure of PredictabilityROC: performance compares predictive capability with actual outcome.Sensitivity vs 1-SpecificityArea under ROC (AUROC)

Value of 1.0 = Perfect resultValue of 0.5 = Toss of a coin

Page 8: PROSTATE CANCER: The Diagnostic  · PDF file1 PROSTATE CANCER: The Diagnostic Dilemma ... zInputs: PSA/VOL/DRE/TRUS/BX HX ... Djavan B 272 ANN/LR Eastham JA 700 LR Age,

8

Predictive Models

0.040.77Artificial Neural Network

0.040.76Logistic Regression

SDAUROC (MEAN)MODEL

Predictive Models

Logistic Regression (LR) Artificial Neural Networks (ANN)

Equivalent good performanceRobust modelsAUROC= 0.76

Porter CR et al, Urology, 60, 5, 831-4, 2002

Predictive Models

Both LR and ANN(+) bx vs (–) bx randomly selected76% of men with (+) bx would have a higher predicted probability of (+) bx

Porter CR et al, Urology, 60, 5, 831-4, 2002

Benefits of Neural Nets

No arbitrary method of variable reductionNo decrease in degree of freedomComplex variable interactions capturedAmerican Joint Committee on Cancer : Endorsed Neural Nets for PCA

– Bostwick D, Semin. Urol Onc. 17,4,1999

Predictive Models:An International Prospective Multicenter Model Involving

4,788 Men

Virginia Mason, WashingtonStanford University, CaliforniaUniversity of Innsbruck, Austria

Presented at ASCO, 2004

Methods

Tyrol, Austria (n=3814)Virginia Mason (n=491)Stanford (n=483)

Page 9: PROSTATE CANCER: The Diagnostic  · PDF file1 PROSTATE CANCER: The Diagnostic Dilemma ... zInputs: PSA/VOL/DRE/TRUS/BX HX ... Djavan B 272 ANN/LR Eastham JA 700 LR Age,

9

Tyrol

Screening StudyElevated PSA: Age-specificRepeat PSADRETRUS TRUS Biopsy (Sextant)

Reissigl A, et al, Cancer 80: 1818-29, 1997

VMMC & Stanford University

Referral populationElevated PSA Abnormal DRETRUS Biopsy (10 core)

Methods

Input variablesAgePSAGland volumePSADDRETRUS

Methods

Tyrol modelData split randomly into three 1200 setsThree ANN modelsThree logistic regression models Cross-way internal validation The median ANN model selectedThe median LR model selected

Methods

Extramural validationMedian ANN & LR models selectedValidated individually against VMMC and SU data sets

Page 10: PROSTATE CANCER: The Diagnostic  · PDF file1 PROSTATE CANCER: The Diagnostic Dilemma ... zInputs: PSA/VOL/DRE/TRUS/BX HX ... Djavan B 272 ANN/LR Eastham JA 700 LR Age,

10

0.3-1,300Yes0.75/0.75

Age, PSA, volume, DRE, TRUS, prior biopsy

ANN/LR4,788Porter CR

0-389Yes0.77/0.76Age, PSA, volume, DRE, TRUS, race, prior biopsy

ANN/LR319Porter CR

4-10Yes0.91/0.90PSA, PSAD, free PSA, PSA TX, DREANN/LR974Djavan B

2.5-4Yes0.87/0.85PSA, PSAD, free PSA, PSA TZ, DREANN/LR272Djavan B

0-4Yes0.75Age, race, PSALR700Eastham JA

2.5-4Yes0.74Age, creatinine kinase PAP, PSA, free PSA

ANN151Babian RJ

>4YesN/AAge, PSA, DRE, TRUSANN1.787Snow PB

PSArangeValidatedAccuracy

ROCInput variablesModelnAuthor

Critique

Study populationsScreening (Model)Referral (Validation)

BiopsySextant vs 10-core

Predictive Models

Accuracy in predicting the outcome of the biopsy75% overall

Porter CR et al., Proceedings of ASCO, 2004

Now we know

Who…

But WHERE??!!

A Patient’s Question

“So, let me get this straight…patients with PSA 2-10 have about a 25%

chance of positive biopsy…

…That means 75% have a chance of negative biopsy.”

A Difficult Sell

“Well Sir, we can’t see the cancer,so we will take more biopsies –

say 10-12…

If we miss it this time we’ll need to do it again, because we miss about 1 in 3.”

Page 11: PROSTATE CANCER: The Diagnostic  · PDF file1 PROSTATE CANCER: The Diagnostic Dilemma ... zInputs: PSA/VOL/DRE/TRUS/BX HX ... Djavan B 272 ANN/LR Eastham JA 700 LR Age,

11

Grey Scale Ultrasound

InaccurateFew targets30-50% false negativeWeather map

Spectral Analysis for Detecting and Evaluating

Prostate CancerErnest J. Feleppa, Andrew Kalisz, SK Alam, Stella Urban

Riverside Research Institute, New York, NYChristopher R. Porter

Virginia Mason Medical Center, Seattle, WAJohn Gillespie

National Institutes of Health, Bethesda, MDPeter Scardino, Makato Ohori

Memorial Sloan-Kettering Cancer Center, New York, NYsupported by:NIH grant CA53561

Prostate Studies: Rationale

MotivationsInability to image prostate cancerInability to target treatment

ObjectivesTo characterize cancerous vs. noncancerous tissuesTo image cancerous regions of the prostate

Prostate Studies: Rationale

To more effectively:Guide biopsies Treat cancerous tissues Spare non-cancerous tissuesMonitor treatment

Page 12: PROSTATE CANCER: The Diagnostic  · PDF file1 PROSTATE CANCER: The Diagnostic Dilemma ... zInputs: PSA/VOL/DRE/TRUS/BX HX ... Djavan B 272 ANN/LR Eastham JA 700 LR Age,

12

SCANNER

COMPUTERSYSTEM

probeand

biopsyneedle

RFechoes

timing andcontrolsignals

SYSTEM FOR ACQUIRING U/S DATA:-- RF echo-signal data (3-D and Bx)-- instrument settings (gain, etc.)-- B-mode-interpretation LOS values

GOLD STANDARD:-- histology biopsy

Spectral Analysis: The Basics

Returning sound wave capturedInstead of using volume (Grey scale/B mode)Entire radio frequency signal analyzedSlope, y intercept and mid band RF values Machine Learning Mathematical Model (ANN)ANN weights the RF values (Look-up table)New RF values analyzed and the probability of the region of interest harboring CAP is calculated

Processing GUIROI and Spectrum

(basis for characterizing tissue)

Power Spectrum: Tissue Properties

Slope = f (size, attenuation)Intercept = f (size, concentration, relative impedance)Midband = f (size, concentration, relative impedance)

General Approach: 3 Steps

1. Build data base

2. Classify tissue types

3. Generate tissue-type images

Page 13: PROSTATE CANCER: The Diagnostic  · PDF file1 PROSTATE CANCER: The Diagnostic Dilemma ... zInputs: PSA/VOL/DRE/TRUS/BX HX ... Djavan B 272 ANN/LR Eastham JA 700 LR Age,

13

General Approach 1: Data Base

Compute spectral

parameter values of RF

data in known location (e.g., biopsy site)

Obtain “truth”data from

histology (e.g., pathology report)

Obtain “general”data for case (e.g., age, race, or PSA)

DATA BASE

Obtain “baseline”data (e.g., LOS) to

assess efficacy

Clinical Data-acquisition System

0

0.2

0.4

0.6

0.8

1

0 0.2 0.4 0.6 0.8 1

best neural net curve: area = 0.848 ± 0.047

average neural net curve: area = 0.804 ± 0.052

B-mode-based curve: area = 0.662 ± 0.034

TPF

FPF

Early ROC Curves for 1005 Biopsies ROC Curves

0.7213

0.4580

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

FPF (1 - specificity)

TPF

(sen

sitiv

ity)

RBF = 0.810 +/- 0.02595% confidence = 0.757 to 0.855

LOS Area = 0.647 +/- 0.02995% confidence = 0.589 to 0.702

Accuracy of Spectral Analysis in Identifying CaP

0.66 +/- 0.030.80 +/- 0.05MSKCC

0.66=/- 0.03O.81 +/-0.05MSKCC

0.67 +/- 0.040.78 +/- 0.05SUNY

AccuracyGrey scale

AccuracySpect. Anal

Site

3-D Multi-Modality Studies

U/S + MRS (+ MRI ) in 3-DIndependent information, e.g., MRS choline/citrate ratio

Radical-prostatectomy patientsData acquired in vivo and in vitro

Correlate with histology in 3-DRegion-by-region (volumetric)

Develop more-powerful classifiersLinear (?)Nearest neighbor (?)Neural network (?)

Page 14: PROSTATE CANCER: The Diagnostic  · PDF file1 PROSTATE CANCER: The Diagnostic Dilemma ... zInputs: PSA/VOL/DRE/TRUS/BX HX ... Djavan B 272 ANN/LR Eastham JA 700 LR Age,

14

3-D Spectral Analysis

Men prior to prostatectomy undergo TRUS using Spectral Analysis

0.2 cm Spectral Analysis cuts are obtained

RRP specimen pathology compared to pre-op Spectral Analysis scans

Anterior Tumor:2-D Images and Histology

(apical view)

Targeting: Anterior Tumor

ppm 2.0 3.5 3.0 2.5

CholineCreatine

Citrate

Cancer2.0

1.5ppm3.5 3.0 2.5

Choline

CreatineCitrate

Healthy

Prostate Cancer

Decreased Citrateloss of cellular function

loss of ductal morphology

Increased Choline + Creatineincreased proliferationmembrane changes

increased cell density

Role of MRS

John Kurhanewicz, Ph.D.MSRC, UCSF

OrlaidCholine Image +

Citrate Image

OverlaidCholine/Citrate

image

Cancer

NormalOverlaidCitrateCholineImages

MRS Displays

In 47 patients who had sextant biopsies and combined MRI/MRSI prior to radical prostatectomy and step-section histopathology, we demonstrated significantly (p< 0.05) improved* localization to a sextant of the prostate (i.e. left/right base, midgland and apex) when combining the information.

Tumor Localization using MRI/MRSI and Sextant Biopsy

Sensitivity SpecificityBiopsy Alone 50% 81%MRI (+) + MRSI (+, > 2 s.d.) 56% 82%MRI (+) or MRSI (+, > 2 s.d.) 88% 40%MRI (+) + MRSI (+, > 3 s.d.) 41% 91%MRI (+) or MRSI (+, > 3 s.d.) 78% 52%MRI (+) + MRSI (+) + Biopsy (+) 34% 98%*MRI (+) or MRSI (+) or Biopsy (+) 94%* 38%

Page 15: PROSTATE CANCER: The Diagnostic  · PDF file1 PROSTATE CANCER: The Diagnostic Dilemma ... zInputs: PSA/VOL/DRE/TRUS/BX HX ... Djavan B 272 ANN/LR Eastham JA 700 LR Age,

15

ROC Curves: MRS vs. MRI

sens

itivi

ty

1-specificity0.2 0.4 0.6 0.8 1.0

0.2

0.4

0.6

0.8

1.0

Radiology 1999;213:481-488

MRI+MRSI, Az=0.83

MRI, Az=0.77

Conclusion

Tissue-type images (TTIs) show promisebiopsy guidance and planning (real time)therapy targeting (escalation and sparing, w/ MRS)monitoring therapy (potentially, w/ MRS)

Conclusion

Payoff for biopsy is largefewer missed cancersfewer unnecessary biopsies

Payoff for therapy targeting is largereduced deleterious effects on non-cancersenhanced desired effects on cancers

We’ll know who to shoot and when to shoot’em