Computed Tomographic Virtual Colonoscopy Computer-Aided Polyp Detection in a Screening Population

CP

RI*SA

BmflwcpsmtdcltbRscr1cBpdutncTCp>

c

Csn

GASTROENTEROLOGY 2005;129:1832–1844

omputed Tomographic Virtual Colonoscopy Computer-Aidedolyp Detection in a Screening Population

ONALD M. SUMMERS,* JIANHUA YAO,* PERRY J. PICKHARDT,‡,§ MAREK FRANASZEK,*NGMAR BITTER,* DANIEL BRICKMAN,* VAMSI KRISHNA,* and J. RICHARD CHOI‡,¶

Diagnostic Radiology Department, Warren Grant Magnuson Clinical Center, National Institutes of Health, Bethesda, Maryland; ‡Uniformedervices University of the Health Sciences, Bethesda, Maryland; §National Naval Medical Center, Bethesda, Maryland; and ¶Walter Reed

tspb1ct

cpapsis

pstss1ptap

bppmsm

tr

See editorial on page 2103.

ackground & Aims: The sensitivity of computed to-ographic (CT) virtual colonoscopy (CT colonography)

or detecting polyps varies widely in recently reportedarge clinical trials. Our objective was to determinehether a computer program is as sensitive as opticalolonoscopy for the detection of adenomatous colonicolyps on CT virtual colonoscopy. Methods: The dataet was a cohort of 1186 screening patients at 3edical centers. All patients underwent same-day vir-

ual and optical colonoscopy. Our enhanced gold stan-ard combined segmental unblinded optical colonos-opy and retrospective identification of precise polypocations. The data were randomized into separateraining (n � 394) and test (n � 792) sets for analysisy a computer-aided polyp detection (CAD) program.esults: For the test set, per-polyp and per-patientensitivities for CAD were both 89.3% (25/28; 95%onfidence interval, 71.8%–97.7%) for detecting ret-ospectively identifiable adenomatous polyps at least

cm in size. The false-positive rate was 2.1 (95%onfidence interval, 2.0–2.2) false polyps per patient.oth carcinomas were detected by CAD at a false-ositive rate of 0.7 per patient; only 1 of 2 wasetected by optical colonoscopy before segmentalnblinding. At both 8-mm and 10-mm adenoma sizehresholds, the per-patient sensitivities of CAD wereot significantly different from those of opticalolonoscopy before segmental unblinding. Conclusions:he per-patient sensitivity of CT virtual colonoscopyAD in an asymptomatic screening population is com-arable to that of optical colonoscopy for adenomas8 mm and is generalizable to new CT virtual colonos-

opy data.

olorectal cancer is the second leading cause of cancerdeath in Americans.1 It is known that, with proper

creening, colorectal cancer can be prevented. Unfortu-

he perceived inconvenience and discomfort of existingcreening tests. Virtual colonoscopy (also known as com-uted tomographic [CT] colonography), a CT scan–ased imaging method, has been under study for the past0 years and shows promise as a method of colorectalancer screening that may be acceptable to many pa-ients.2,3

Recent large clinical trials have suggested that virtualolonoscopy may have high sensitivity and specificity forolyp detection.4,5 Other studies have raised questionsbout its reproducibility and accuracy in actual clinicalractice.6–9 If virtual colonoscopy is to be widely dis-eminated for colorectal cancer screening, methods thatmprove consistency and accuracy would be highly de-irable.

Computer-aided polyp detection (CAD) has been pro-osed by a number of investigators to improve the con-istency and sensitivity of virtual colonoscopy interpre-ation and reduce interpretation burden.10 Preliminarytudies of prototype CAD systems on small patient dataets have reported per-polyp sensitivities from 64% to00% and false-positive rates from 1 to 11 false positiveser patient for detecting polyps �1 cm.11–17 However,here is currently insufficient evidence whether CAD isccurate in a screening population and whether the re-orted results generalize to independent data.The purpose of this study was to provide this evidence

y assessing CAD performance on a large, consecutive,rospectively enrolled asymptomatic screening patientopulation. To ascertain the generalizability of perfor-ance of CAD, we randomized the patients’ data into

eparate training and test sets and evaluated the perfor-ance of CAD on each data set.

Abbreviations used in this paper: CAD, computer-aided polyp detec-ion; CI, confidence interval; CT, computed tomographic; FROC, free-esponse receiver operating characteristic.

© 2005 by the American Gastroenterological Association0016-5085/05/$30.00

doi:10.1053/j.gastro.2005.08.054

a(dpcgroitp

c1md

madsHrma

ribfmmo(

eTrp

mt

3ovmtv

mct(sbg

cbsmprpPrd

wcctbp

pcc

gddtdfp

lt

December 2005 COMPUTER–AIDED POLYP DETECTION 1833

Patients and Methods

Patient Population

The patient population consisted of 1253 asymptom-tic adults between 40 and 79 years of age at 3 medical centersinstitutions 1–3), of whom 1233 underwent complete same-ay virtual and optical colonoscopy.4 Twenty of the 1253atients were excluded because of incomplete optical colonos-opy, inadequate preparation, or failure of the CT colono-raphic system. The study was approved by the institutionaleview boards at all 3 centers. Written informed consent wasbtained from all patients. This study was part of the originalnstitutional review board–approved project and consent formhat led to publication of the study by Pickhardt et al,4 and theatient population is the same.

Bowel Preparation

Patients underwent a 24-hour colonic preparation thatonsisted of oral administration of 90 mL sodium phosphate,0 mg bisacodyl, 500 mL barium (2.1% by weight), and 120L diatrizoate meglumine and diatrizoate sodium given in

ivided doses.18

CT Scanning

A small, flexible rectal catheter was inserted and pneu-ocolon achieved by patient-controlled insufflation of room

ir. Each patient was scanned in the supine and prone positionsuring a single breath hold using a 4-channel or 8-channel CTcanner (General Electric LightSpeed or LightSpeed Ultra; GEealthcare Technologies, Waukesha, WI). CT scanning pa-

ameters included 1.25- to 2.5-mm section collimation, 15m/s table speed, 1-mm reconstruction interval, 100 mAs,

nd 120 kVp.

Optical Colonoscopy

Optical colonoscopy was performed by 1 of 17 expe-ienced colonoscopists. Our technique for segmental unblind-ng of virtual colonoscopy results at optical colonoscopy haseen previously described4 and reduces optical colonoscopyalse negatives as much as 12% for large adenomas (�10

m).19 The colonoscopists used a calibrated guidewire toeasure polyp size and recorded whether the polyp was located

n a haustral fold and a subjective assessment of polyp shapesessile, pedunculated, or flat).

CT Colonography Database

CT images from the virtual colonoscopy studies fromach of the 3 institutions were loaded onto a computer server.he CT images from 47 patients could not be located or

estored and were excluded from further analysis; this left 1186atients with complete data.

Recording the ground truth. To assess the perfor-ance of the CAD software, we developed an enhanced ground

ruth (calibration data) based on manual determination of the p

-dimensional borders of polyps. Each polyp �6 mm found atptical colonoscopy was located on the prone and supineirtual colonoscopy examinations using 3-dimensional endolu-inal reconstructions with “fly-through” capability and mul-

iplanar reformatted images (Viatronix V3D colon, researchersion 1.3.0.0; Viatronix, Stony Brook, NY).

For each polyp and for each position (supine and prone), aarker was placed manually in the center of each polyp using

omputer software. Then the borders of the polyp on each slicehat contained the polyp were manually traced. The markersapproximately 500) and borders (approximately 3650) weretored in data files. The markings and tracings were performedy a trained research assistant (D.B.) supervised by a radiolo-ist (R.M.S.).

Radiologist false positives. To assess the potentiallinical significance of CAD false positives, we created a data-ase of radiologist false positives to enable comparison of the 2ets for any commonality. This database allowed us to deter-ine whether radiologists and CAD made the same false

ositives. A trained research assistant (V.K.), supervised by aadiologist (R.M.S.), identified the false-positive polyps re-orted on the same cases by the radiologists in the study byickhardt et al.4 Each false positive that was identifiable inetrospect was marked and manually traced as previouslyescribed.

CAD System

The CAD system has been described in detail else-here.12,17 It consisted of automated identification of the

olonic lumen and wall,20 electronic subtraction of opacifiedolonic fluid,21 calculation of colonic surface features, segmen-ation of candidate polyps to locate their entire 3-dimensionaloundaries,22 and classification to distinguish true- and false-ositive polyp detections.23,24

The output of the CAD system was a series of locations ofolyp candidates in the CT images. The location data could beonverted to a graphical overlay on 3-dimensional virtualolonoscopy images.

Matching the Ground Truth and ComputerDetections

The CAD software compared its detections with theround truth tracings in a blinded fashion. If any part of aetection matched any part of a manual tracing of a polyp, theetection was considered a true positive; otherwise, the detec-ion was considered a false positive. Similarly, if any part of aetection matched any part of a manual tracing of a radiologistalse positive, the detection was considered a matching falseositive.

Training Method and Testing

As for other types of radiology CAD such as detectingung nodules on CT scans or breast cancer on mammography,he CAD system for detecting polyps must be trained on

htsa

totqtsatCs

vahwopte

ifs

pTec

Rorclas(a8w

iCgypcsndiipcBwt

ts

T

NNNNNA

T

N

NN

Oa

b

1834 SUMMERS ET AL GASTROENTEROLOGY Vol. 129, No. 6

ow to discriminate between true polyps and nonpolyps. Afterraining, the entire CAD system, including the classifier,hould be applied to new “test” cases to provide a fairerssessment of future performance.

To implement this, the data set was divided into separateraining and test sets. We chose to train on one third and testn the remaining two thirds of the data. This partitioning ofhe data enables better statistical power during testing anduicker processing during technical development when theraining set is used. The division into training and test dataets was conducted using a random number generator thatssigned patients from all 3 centers to either the training orest sets (Microsoft Access; Microsoft Corp, Redmond, WA).haracteristics of the patients in the training and test sets are

hown in Table 1.Testing cases were sequestered and not used during de-

elopment or training.25 When an acceptable training wasccomplished, testing was run to produce the results shownerein. We did perform training and testing with andithout merging of overlapping detections; however, basedn superior performance with merging during training, weresent only results for merged detections. Details of theraining and classifier design have been previously report-d.23,24,26

The training was performed using detections from the train-ng set cases from all 3 institutions. Training was performedor adenomas at 10-, 8-, and 6-mm size thresholds. Adenomasmaller than these size thresholds and all nonadenomatous

able 1. Patient Population in the Database

Train(n � 394)

Test(n � 792)

o. of men (%) 227 (57.6) 473 (59.7)o. of women (%) 167 (42.4) 319 (40.3)o. at institution 1 (%) 122 (31.0) 283 (35.7)o. at institution 2 (%) 123 (31.2) 190 (24.0)o. at institution 3 (%) 149 (37.8) 319 (40.3)ge, y (mean � SD) 58.0 � 7.4 57.7 � 7.1

able 2. Polyps Identified

At OC

Train(n � 99)

o. of adenomas (%)6–7 mm 32 (32.3)8–9 mm 18 (18.2)�10 mm 19 (19.2)

o. of carcinomas (%) 0 (0.0)o. of hyperplastic polyps (%)6–7 mm 21 (21.2)8–9 mm 6 (6.1)�10 mm 3 (3.0)

C, optical colonoscopy; VC, virtual colonoscopy.Polyps identified at OC, including those found after unblinding.

olyps were placed in the false-positive set during training.he outputs of the training were 3 different classifiers, one forach size threshold, that were individually applied to the CTolonography test data.

The CAD software executed on both the Linux (Redhat,aleigh, NC) and Microsoft Windows (Microsoft Corp)perating systems. The majority of the cases (�99%) wereun on a Linux supercluster (a network of inexpensiveomputers linked together) to more efficiently analyze thearge number of CT colonography examinations.27 As manys 64 examinations could be analyzed simultaneously on theupercluster. CAD successfully analyzed all but 4 training2 supine and 2 prone) and 3 test examinations (2 supinend 1 prone). The processing time per patient was 20.2 �.0 minutes (n � 1179), approximately half of which timeas spent reading the images across the network.

Data Analysis

We used free-response receiver operating character-stic (FROC) analysis, the standard method for evaluatingAD performance.28 FROC analysis produces curves thatraphically show the sensitivity of CAD for detecting pol-ps versus false-positive rate (number of false positives peratient) for different settings of a tunable parameter in thelassifier. As is typical in CAD, one can tune the CADystem to yield higher sensitivity at the expense of a greaterumber of false positives. FROC curves are presented forifferent adenoma size categories and for training and test-ng. Because we are focusing on the more clinically signif-cant adenomatous polyps, true-positive detections onroven nonadenomatous polyps were ignored and not in-luded in the false-positive rates for the FROC analysis.ecause the number of nonadenomatous polyps (Table 2)as small relative to the number of patients, the effect of

his procedure on false-positive rates is negligible.While FROC curves show the spectrum of CAD sensitivi-

ies across a range of false-positive rates, for clinical use a CADystem is typically set at a specific operating point on the

At retrospective VC interpretationb

Test(n � 204)

Train(n � 79)

Test(n � 173)

82 (40.2) 24 (30.4) 67 (38.7)26 (12.7) 17 (21.5) 24 (13.9)29 (14.2) 19 (24.1) 28 (16.2)2 (1.0) 0 (0.0) 2 (1.2)

34 (16.7) 12 (15.2) 27 (15.6)18 (8.8) 4 (5.1) 16 (9.2)13 (6.4) 3 (3.8) 9 (5.2)

a

FFpfTtsorp

oatwtrp

pasmcwsdp

yarfoausecCc

tIpBrsemdr

asd

d�clswoa

oTpsfso0i

fC

F((apEpT

December 2005 COMPUTER–AIDED POLYP DETECTION 1835

ROC curve with fixed sensitivity and false-positive rate.or each of the 3 size thresholds, we selected an operatingoint on the FROC curve. We report the sensitivities andalse-positive rates at these operating points in the tables.he operating points were chosen in relatively flat parts of

he FROC curves where there were diminishing gains inensitivity as the false-positive rates were increased. Theperating points were chosen somewhat arbitrarily but rep-esent reasonable tradeoffs between sensitivity and false-ositive rates.

Assessments of false positives. A random subsetf 64 false positives was selected from those found afterpplication of the classifier trained on adenomas �10 mmo determine their cause. Images of these false positivesere loaded into a software application developed by one of

he authors (J.Y.) that creates a mosaic of images that can beeviewed rapidly to determine the cause of the falseositives.

Subgroup analyses. To better characterize CADerformance, we computed the sensitivity of CAD 3 ways: forll polyps, for those surrounded by luminal air, and for thoseubmerged in opacified fluid. A polyp was considered sub-erged if by visual assessment �50% of its surface was

overed by fluid. Polyps were not considered submerged if theyere merely coated with a thin layer of opacified fluid. We also

tratified detection performance by polyp shape (sessile, pe-unculated, or flat), location in the colon, and whether theolyps were on folds.

Statistical Analysis

Sensitivity was computed 2 ways: (1) using all pol-ps found at segmentally unblinded optical colonoscopynd (2) using only those polyps visible on retrospectiveeview of the CT colonography images. The former is usefulor comparing the overall sensitivity of CAD with that ofptical colonoscopy before segmental unblinding and liter-ture reports of radiologist interpretation. The latter isseful for distinguishing the performance of CAD fromhortcomings of the CT colonography technique itself. Forxample, some polyps, particularly those 6 or 7 mm in size,ould not be found on the supine and/or prone views.onsequently, it is not possible to train on them or toonfirm whether CAD detected them.

We report exact 95% confidence intervals (CIs) for sensi-ivities and false-positive rates (SAS software version 9.1; SASnstitute Inc, Cary, NC), used the Fisher exact test to compareroportions, and consider statistical significance to be P � .05.ootstrapping was used to compute standard deviations over a

ange of operating points for the FROC analysis. The boot-trapping was conducted by determining FROC curves forach of 100 random samples of 792 test patients with replace-ent (duplicates allowed) and then estimating the standard

eviation at fixed values of the sensitivity and false-positive

Results

The patients were distributed into the trainingnd test sets as shown in Table 1, with similar age andex distributions, accounting for the 2:1 split. The polypistributions are shown in Table 2.The FROC curves are shown in Figure 1 for the 3

ifferent classifiers trained to detect adenomatous polyps10, �8, and �6 mm. These curves indicate that at a

onstant false-positive rate, sensitivity was higher forarger polyps. Sensitivity was also higher on the traininget compared with the test set, although the differencesere small (�5%) for the 8-mm and 10-mm size thresh-lds. The 3 operating points are indicated by theirssociated error bars.

The per-polyp and per-patient sensitivities at theperating point at each size threshold are shown inable 3. At a false-positive rate of 2.1 per patient forolyps �10 mm, the per-polyp and per-patient sen-itivities were both 89.3%. Both carcinomas wereound at a false-positive rate of 0.7 per patient. Theensitivities were lower for the 2 smaller-size thresh-lds. Example virtual colonoscopy images of 1.4-,.8-, and 0.6-cm polyps detected by CAD are shownn Figures 2– 4.

The sensitivities of first-look optical colonoscopy (be-ore segmental unblinding) and virtual colonoscopyAD, using a baseline of all adenomas found by segmen-

igure 1. FROC curves for the training (open symbols) and testclosed symbols) sets are shown for adenomatous polyps �10 mmcircles), �8 mm (squares), and �6 mm (triangles). Pooled data fromll 3 medical centers are shown. We show only the clinically relevantortion where the number of false positives (FP) per patient is �10.rror bars (1 SD) from bootstrap analysis of sensitivity and false-ositive rate are shown at the 3 operating points for the test set fromable 3.

ally unblinded optical colonoscopy, are compared in

Tscptmi

6paioi0

aflwit

lidstppiwd

8iwbr

svpf

aappr

shpthtiplcigr

aso2pypt

T

S

S

F

NpC

1836 SUMMERS ET AL GASTROENTEROLOGY Vol. 129, No. 6

able 4. The per-patient sensitivities of CAD were notignificantly different from that of first-look opticalolonoscopy at the 8-mm and 10-mm size thresholds; theer-polyp sensitivities were not significantly different athe 10-mm size threshold. Optical colonoscopy initiallyissed 1 of the 2 carcinomas before segmental unblind-

ng; CAD detected both cancers.Standard deviations of sensitivity ranged from 4% to

% and of false-positive rate ranged from 0.1 to 0.3 peratient at the operating points (Figure 1). The bootstrapnalysis revealed that the standard deviations in sensitiv-ty increased at lower false-positive rates to a maximumf 10%. The standard deviations in false-positive ratencreased at higher false-positive rates to a maximum of.8 per patient.Sensitivity was higher for adenomatous polyps in the

ir-filled part of the colonic lumen compared with theuid-filled part (Table 5). The sensitivity differencesere statistically significant for 5 of 6 pairwise compar-

sons. In general, polyps were more frequently located inhe air-filled part of the colonic lumen.

Sensitivity of polyp detection as a function of shape,ocation, and relationship to a haustral fold is shownn Table 6. Larger polyps were most frequently pe-unculated, and smaller polyps were most frequentlyessile. For the 6-mm and larger polyps, CAD sensi-ivity was lower for sessile polyps compared withedunculated polyps and for polyps on a fold com-ared with polyps not on folds. There were no signif-cant differences in sensitivity for left-sided comparedith right-sided polyps. None of 5 flat polyps wereetected by CAD.Of CAD false negatives, 67% (2/3), 90% (9/10), and

9% (41/46) were for adenomatous polyps on or touch-ng a fold and 67% (2/3), 80% (8/10), and 24% (11/46)ere on or near (within a few voxels of) the air-fluidoundary at the 10-, 8-, and 6-mm size thresholds,

able 3. Performance Characteristics of Virtual ColonoscopyReview

Ade

�6 mm �

ensitivity according toadenoma

73/119 (61.3% [52.0–70.1]) 42/52 (80.

ensitivity according topatient

72/95 (75.8% [65.9–84.0]) 41/47 (87.

alse positives perpatient

7.9 (7.7–8.1) 6.7 (6.5

OTE. Sensitivities for detection of adenomatous polyps in the teolyps found on retrospective review of virtual colonoscopy imagesI).

espectively. i

Analysis of 64 random CAD false positives �1 cmhowed that the majority were caused by the ileocecalalve (52/64; 81%) at a false-positive rate of 2.1 peratient. The remainder was due to haustral or rectalolds, residual stool or fluid, or other causes.

The radiologists identified 165 false-positive polyps ofll sizes in the test set, of which 126 could be found ont least 1 view (supine or prone). Of 1692 CAD false-ositive detections in the test set (false-positive rate, 2.1er patient), only 15 CAD false positives (0.9%) matchedadiologist false positives.

Discussion

CT virtual colonoscopy has progressed rapidlyince its inception in 1994.29 Several large clinical trialsave been reported.4,6,8,30 Some of these trials have re-orted excellent sensitivity, but others have shown rela-ively poor sensitivity. The causes of poor sensitivitiesave been variously attributed to out-of-date CT scannerechnology, absence of bowel opacification, inadequatenterpretation software, improper interpretation ap-roach (2-dimensional rather than 3-dimensional), orack of training of the interpreters.7,31–34 While there isonsensus that virtual colonoscopy is appropriate forndications such as incomplete colonoscopy, there is on-oing debate about its role in the asymptomatic average-isk (screening) patient.

The process of interpreting virtual colonoscopy ex-minations is an area that has received considerablecrutiny in recent years. For example, there is debatever whether images should be read using a primary-dimensional versus primary 3-dimensional ap-roach, whether different interpretation softwareields different results, and whether training or occu-ation affect interpretation skill.6,7,9,35–39 It is clearhat different observers interpret virtual colonoscopy

for the Detection of Adenomas Based on Retrospective

s

�10 mm Carcinomas

7.5–90.4]) 25/28 (89.3% [71.8–97.7]) 2/2 (100% [15.8–100])

4.3–95.2]) 25/28 (89.3% [71.8–97.7]) 2/2 (100% [15.8–100])

2.1 (2.0–2.2) 0.7 (0.6–0.8)

t are expressed as number/total number (% [95% CI]) based one-positive rates per patient are expressed as mean number (95%

CAD

noma

8 mm

8% [6

2% [7

–6.9)

st se. Fals

mages with different levels of skill. For example,

F�Bstpetp

bC

lsctfwhlpts

December 2005 COMPUTER–AIDED POLYP DETECTION 1837

letcher et al found that 17 of 30 false-negative polyps1 cm were missed because of perceptual error.40

y detecting disease on radiologic images with highensitivity and low false-positive rate, CAD can po-entially improve overall physician interpretativeerformance, diminish the frequency of perceptualrrors, and allow more poorly performing interpreterso attain performance levels comparable to ex-erts.41,42

A number of CAD systems for polyp detection haveeen described.12,14,43–52 In a typical implementation,

ike shapes that protrude into the colonic lumen. Factorsuch as colonic wall thickness, surface curvature, andontrast enhancement have been proposed as useful fea-ures that can be quantitated and can distinguish polypsrom normal colonic mucosa.11–14,17,44,47,53 While theseorks are encouraging, in general they have used smallighly selected patient populations, unclear patient se-ection criteria, or more readily detectable conspicuousolyps to develop and assess the CAD system. In addi-ion, with few exceptions,54,55 data have come from aingle institution with testing performed on the same

gure 2. (A) Optical and (B and C) 3-dimensional virtual colonos-py images of a 1.4-cm polyp in the transverse colon of a 64-year-

woman in the test set. The blue coloring in C indicates the partthe polyp detected by CAD. A portion of the colon centerline isown in green in B and C.

Ficooldofsh

ata used for training.

pepdtdecaa

csddtcslsc

1838 SUMMERS ET AL GASTROENTEROLOGY Vol. 129, No. 6

While CAD development for polyp detection hasroceeded along many fronts, a common and criticallement is validation of performance on a database ofroven cases. There are many important issues abouteveloping the database and validating performance ifhe CAD system is to be generalizable to new patientata. It is accepted by many experts that the keylements of the database are that it be an unbiasedollection of proven cases of sufficient number todequately reflect the diversity of polyp sizes, shapes,

ritical to determine the generalizability of the CADystem by assessing its performance on a fresh set ofata (a test set) different from that on which it waseveloped (the training set). Our database and valida-ion methods were chosen to fulfill these importantriteria. In this study, we used data from 1253 con-ecutive screening cases from 3 medical institutions,ess about 5% that were excluded, and divided it intoeparate training and testing samples. The CTolonography data were validated with an enhanced

gure 3. (A) Optical and (B and C) 3-dimensional virtual colonoscopyages of a 0.8-cm polyp in the sigmoid colon of a 60-year-old man ine test set. The blue coloring in C indicates the part of the polyptected by CAD. A portion of the colon centerline is shown in green inand C.

FiimthdeB

old standard: segmentally unblinded optical colonos-

cc

opr8sfTsc

clt1Co[g[[

December 2005 COMPUTER–AIDED POLYP DETECTION 1839

opy. To our knowledge, this is the largest virtualolonoscopy database of its kind.

When we analyzed all polyps visible in retrospectn CT colonography, both the per-polyp and per-atient sensitivities were 89.3%, at a false-positiveate of 2.1 per patient for polyps �10 mm. At the-mm size threshold, the per-polyp and per-patientensitivities were 80.8% and 87.2%, respectively, at aalse-positive rate of 6.7 false polyps per patient.hese results indicate that CAD reliably finds retro-

pectively visible adenomatous polyps �8 mm on CT

When compared with sensitivities of first-look opticalolonoscopy and with radiologist interpretation in theargest CT colonography trials, the per-adenoma sensi-ivity (86.2%) of CAD was equivalent or better at the0-mm size threshold. For example, the sensitivity ofAD was not significantly different compared with thatf radiologists, as reported by Pickhardt et al (47/5192.2%]; 95% CI, 81.1–97.8), but was significantlyreater than that reported by Cotton et al (28/5452.0%]; 95% CI, 38.7–65.3), Rockey et al (35/5564%]; 95% CI, 49–77), and Johnson et al (double read;

e 4. (A) Optical and (B and C) 3-dimensional virtual colonos-images of a 0.6-cm polyp in the transverse colon of a 65-year-an in the test set. The blue coloring in C indicates the part ofolyp detected by CAD. A portion of the colon centerline isn in green in B and C.

Figurcopyold mthe pshow

6/41 [63.4%]; 95% CI, 46.9–77.9).4,6–8 Note that

Cphcc

lrraafeot(i(((4aslAitCaa

gsitc

t8Ssraft

ttefawCaCscpcf

ardwcffifps

T

S

S

Ncoa ).

1840 SUMMERS ET AL GASTROENTEROLOGY Vol. 129, No. 6

otton et al did not break down per-polyp sensitivity byolyp histology so that all colorectal lesions (includingyperplastic polyps) were included. Rockey et al reportedombined sensitivities for detecting adenomas and can-ers.

Similarly, when compared with sensitivities of first-ook optical colonoscopy (85.7% and 89.6%) and withadiologist interpretation in the largest CT colonog-aphy trials, per-patient sensitivities for CAD (89.3%nd 85.4%) were equivalent or better at the 10-mmnd 8-mm size thresholds, respectively, and are there-ore likely to be in the clinically acceptable range. Forxample, at the 10-mm size threshold, the sensitivityf CAD was not significantly different compared withhat of radiologists, as reported by Pickhardt et al45/48 [93.8%]; 95% CI, 82.8 –98.7), but was signif-cantly greater than that reported by Cotton et al23/42 [55.0%]; 95% CI, 39.9 –70.0), Rockey et al37/63 [58.7%]; 95% CI, 45–71), and Johnson et aldouble read; 30/47 [63.8%]; 95% CI,8.5–77.3).4,6 – 8 Note that Cotton et al, Rockey et al,nd Johnson et al did not break down per-patientensitivity by polyp histology so that all colorectalesions (including hyperplastic polyps) are included.t the 8-mm size threshold, our per-patient sensitiv-

ties were not significantly different compared withhat reported by Pickhardt et al (77/82 [93.9%]; 95%I, 86.3–98.0). These comparisons do not take intoccount any changes in specificity that might occur asconsequence of CAD false positives.We found that CAD developed on training data was

eneralizable to a separate test set. For example, theensitivity and false-positive rate of CAD were essentiallydentical for the training and test sets at the 10-mm sizehreshold. For smaller size thresholds, there was a de-

able 4. Performance Characteristics of Virtual ColonoscopyAdenomas Based on All Adenomas

Ade

�6 mm �

ensitivity according toadenoma

CAD 73/137 (53.3% [44.6–61.9])a 42/55 (76.Optical colonoscopy 122/137 (89.1% [82.6–93.7])a 50/55 (90.

ensitivity according topatient

CAD 72/109 (66.1% [56.4–74.9])c 41/48 (85.Optical colonoscopy 95/109 (87.2% [79.4–92.8])c 43/48 (89.

OTE. Sensitivities for detection of adenomatous polyps in the teolonoscopy CAD and at first-look (before segmental unblinding) optiptical colonoscopy.–cP � .05 for pairwise comparison of sensitivities (Fisher exact test

rease in sensitivities between the training and test sets f

hat ranged from about 5% to 10% on average at the-mm and 6-mm size thresholds, respectively (Figure 1).tandard deviations at the operating points were low forensitivity (4%–6%) and negligible for false-positiveate (0.1–0.3). These standard deviations, which providen estimate of the expected change in sensitivities andalse-positive rates on new data sets, are likely to be inhe clinically acceptable range.

For guiding practical use by clinicians and futureechnical improvements by researchers, it is importanto ascertain particular situations in which CAD is lessffective. The sensitivity of our CAD system was loweror polyps under fluid, for small sessile and flat polyps,nd for small polyps on folds. Many false negativesere at the air-fluid boundary, a location difficult forAD to analyze. Factors such as the CT attenuationnd amount of opacified colonic fluid may also affectAD performance. The bowel preparation used in this

tudy produced a relatively large volume of residualolonic fluid.56 Subsequent modifications of the bowelreparation have since reduced the amount of retainedolonic fluid, which would likely improve CAD per-ormance.

The significance of the false-positive rate is harder tossess. Physician acceptance of 2.1 or 6.7 false-positiveates, at the 10-mm and 8-mm thresholds, respectively,epends on a number of issues: the efficiency (speed) withhich physicians can review CAD “hits” and how diffi-

ult it is to decide if a CAD hit is true or false. Theormer is determined by the quality of the user interfaceor the interpretation software and was not specificallynvestigated by us. The latter was studied by us at aalse-positive rate of 2.1. We found that most falseositives were readily identified to be normal structuresuch as the ileocecal valve or colonic folds. In addition,

and First-Look Optical Colonoscopy for the Detection of

s

m �10 mm Carcinomas

3.0–86.8])b 25/29 (86.2% [68.3–96.1]) 2/2 (100% [15.8–100])0.0–97.0])b 25/29 (86.2% [68.3–96.1]) 1/2 (50.0% [1.3–98.7])

2.2–93.9]) 25/28 (89.3% [71.8–97.7]) 2/2 (100% [15.8–100])7.3–96.5]) 24/28 (85.7% [67.3–96.0]) 1/2 (50.0% [1.3–98.7])

t are expressed as number/total number (% [95% CI]) at virtuallonoscopy based on all adenomas found on segmentally unblinded

CAD

noma

8 m

4% [69% [8

4% [76% [7

st secal co

ew (0.9%) of the CAD false positives coincided with

rfbic3

sppoetm

ictncAapa

Fhppi

pdCmbya

trCtiii

puabl

e5

.C

ADS

ensi

tivity

for

Aden

omas

Sur

roun

ded

byAi

ror

Flui

d

Aden

omas

�6

mm

Aden

omas

�8

mm

Aden

omas

�10

mm

Sup

ine

Pron

eaS

upin

eaPr

onea

Sup

inea

Pron

ea

47

/89

(52

.8%

[41

.9–6

3.5

])4

7/9

0(5

2.2

%[4

1.4

–62.9

])24/3

2(7

5.0

%[5

6.6

–88.5

])30/3

9(7

6.9

%[6

0.7

–88.9

])13/1

4(9

2.9

%[6

6.1

–99.8

])18/2

2(8

1.8

%[5

9.7

–94.8

])id

9/2

3(3

9.1

%[1

9.7

–61

.5])

5/2

2(2

2.7

%[7

.8–4

5.4

])8/1

9(4

2.1

%[2

0.3

–66.5

])5/1

3(3

8.5

%[1

3.9

–68.4

])6/1

4(4

2.9

%[1

7.7

–71.1

])2/6

(33.3

%4.3

–77.7

))

TE.

Sen

sitiv

ities

for

dete

ctio

nof

aden

omat

ous

poly

psin

the

test

set

are

expr

esse

das

num

ber/

tota

lnu

mbe

r(%

[95%

CI])

base

don

aden

omas

foun

don

retr

ospe

ctiv

ere

view

ofvi

rtua

lon

osco

pyim

ages

.�

.05

for

pairw

ise

com

paris

ons

ofse

nsiti

vitie

sin

each

colu

mn

for

poly

pssu

rrou

nded

byai

rve

rsus

fluid

(Fis

her

exac

tte

st).

December 2005 COMPUTER–AIDED POLYP DETECTION 1841

adiologist false positives. This suggests that most CADalse positives would be rejected by the radiologist aseing unlikely to represent true polyps. There is prelim-nary evidence that CAD false positives do not signifi-antly impair radiologists’ specificity even when almost0 false positives are shown per patient.52

Because of the large number of CT colonography dataets in this study, we used a Linux supercluster toerform the CAD analyses more efficiently. In clinicalractice, the CAD system described herein would be runn a readily available desktop personal computer runningither the Linux or Microsoft Windows operating sys-ems. We estimate the typical processing time to be �10inutes per patient using such a system.This study has several limitations. First, we could have

ncorrectly matched polyps found at optical and virtualolonoscopy. This error could either increase or decreasehe measured sensitivity of CAD. Second, there were aumber of polyps found at optical colonoscopy that weould not find retrospectively at virtual colonoscopy.lthough it is possible that CAD “false positives” were

ctually true-positive detections of such polyps, we sus-ect this occurred infrequently. To avoid bias, we did notttempt to reclassify such polyps.

We do not report performance on hyperplastic polyps.or polyps in the test set �6 mm, 31.9% (65/204) wereyperplastic polyps. While hyperplastic polyps may ap-ear indistinguishable from adenomas on CT colonogra-hy, they have no malignant potential and consequentlyt is less important to detect them.

CT colonography CAD is an active area of researchursued by a number of investigators both in the aca-emic and commercial sectors. Future improvements inAD algorithms will likely lead to even better perfor-ance. CAD systems for CT colonography are likely to

ecome commercially available within the next fewears, pending approval by the appropriate regulatorygencies.

The economics of CT colonography CAD is an impor-ant and open issue. Unlike the situation for mammog-aphy CAD, colonography CAD is not yet reimbursable.AD could decrease expensive radiologist interpretation

ime and missed cancer diagnoses, leading to cost sav-ngs. However, the workup of radiologist false positivesnduced by CAD could increase costs. Each of thesessues will need to be assessed.

In conclusion, we found that the sensitivity and false-ositive rate of CAD in an asymptomatic screening pop-lation were in the range likely to be clinically accept-ble and were generalizable to fresh CT virtual

colonoscopy data.Ta Air

Flu

NO

col

a P

1

1

1

1

1

1

1

1

1

T

A

R

C

NfstLa ).

1842 SUMMERS ET AL GASTROENTEROLOGY Vol. 129, No. 6

References1. Jemal A, Tiwari RC, Murray T, Ghafoor A, Samuels A, Ward E,

Feuer EJ, Thun MJ. Cancer statistics, 2004. CA Cancer J Clin2004;54:8–29.

2. Gluecker TM, Johnson CD, Harmsen WS, Offord KP, Harris AM,Wilson LA, Ahlquist DA. Colorectal cancer screening with CTcolonography, colonoscopy, and double-contrast barium enemaexamination: prospective assessment of patient perceptions andpreferences. Radiology 2003;227:378–384.

3. van Gelder RE, Birnie E, Florie J, Schutter MP, Bartelsman JF,Snel P, Lameris JS, Bonsel GJ, Stoker J. CT colonography andcolonoscopy: assessment of patient preference in a 5-week fol-low-up study. Radiology 2004;233:328–337.

4. Pickhardt PJ, Choi JR, Hwang I, Butler JA, Puckett ML, HildebrandtHA, Wong RK, Nugent PA, Mysliwiec PA, Schindler WR. Computedtomographic virtual colonoscopy to screen for colorectal neopla-sia in asymptomatic adults. N Engl J Med 2003;349:2191–2200.

5. Pineau BC, Paskett ED, Chen GJ, Espeland MA, Phillips K, Han JP,Mikulaninec C, Vining DJ. Virtual colonoscopy using oral contrastcompared with colonoscopy for the detection of patients withcolorectal polyps. Gastroenterology 2003;125:304–310.

6. Cotton PB, Durkalski VL, Benoit PC, Palesch YY, Mauldin PD,Hoffman B, Vining DJ, Small WC, Affronti J, Rex D, Kopecky KK,Ackerman S, Burdick JS, Brewington C, Turner MA, Zfass A,Wright AR, Iyer RB, Lynch P, Sivak MV, Butler H. Computedtomographic colonography (virtual colonoscopy): a multicentercomparison with standard colonoscopy for detection of colorectalneoplasia. JAMA 2004;291:1713–1719.

7. Rockey DC, Paulson E, Niedzwiecki D, Davis W, Bosworth HB,Sanders L, Yee J, Henderson J, Hatten P, Burdick S, Sanyal A,Rubin DT, Sterling M, Akerkar G, Bhutani MS, Binmoeller K,Garvie J, Bini EJ, McQuaid K, Foster WL, Thompson WM,Dachman A, Halvorsen R. Analysis of air contrast barium enema,computed tomographic colonography, and colonoscopy: prospec-tive comparison. Lancet 2005;365:305–311.

8. Johnson CD, Harmsen WS, Wilson LA, Maccarty RL, Welch TJ,Ilstrup DM, Ahlquist DA. Prospective blinded evaluation of com-puted tomographic colonography for screen detection of colorec-tal polyps. Gastroenterology 2003;125:311–319.

9. Johnson CD, Toledano AY, Herman BA, Dachman AH, McFarland

able 6. CAD Sensitivity According to Adenoma Shape, Relat

�6 mm

denoma shapeSessile 33/68 (48.5% [36.2–61.0])a–c

Flat 0/5 (0.0% [0.0–47.8])b,d

Pedunculated 24/34 (70.6% [52.5–84.9])a,c,d

elationship to haustral foldOn a fold 28/55 (50.9% [37.1–64.7])e

Not on a fold 43/64 (67.2% [54.3–78.4])e

olonic locationLeft colon 40/69 (58.0% [45.5–69.8])Right colon 33/50 (66.0% [51.2–78.8])

OTE. Sensitivities for detection of adenomatous polyps in the test seound on retrospective review. Polyp shape, relationship to a haustrahapes of 12 polyps were described as round, oval, or eccentric. CADhey are shown according to colonic location and whether the polypseft colon, splenic flexure to rectum, inclusive; right colon, cecum to–eP � .05 for pairwise comparison of sensitivities (Fisher exact test

EG, Barish MA, Brink JA, Ernst RD, Fletcher JG, Halvorsen RA Jr,

Hara AK, Hopper KD, Koehler RE, Lu DS, Macari M, Maccarty RL,Miller FH, Morrin M, Paulson EK, Yee J, Zalis M. Computerizedtomographic colonography: performance evaluation in aretrospective multicenter setting. Gastroenterology 2003;125:688–695.

0. Summers RM, Yoshida H. Future directions: computer-aided di-agnosis. In: Dachman AH, ed. Atlas of virtual colonoscopy. NewYork, NY: Springer, 2003:55–62.

1. Yoshida H, Nappi J, MacEneaney P, Rubin DT, Dachman AH.Computer-aided diagnosis scheme for detection of polyps at CTcolonography. Radiographics 2002;22:963–979.

2. Summers RM, Johnson CD, Pusanik LM, Malley JD, Youssef AM,Reed JE. Automated polyp detection at CT colonography: feasi-bility assessment in a human population. Radiology 2001;219:51–59.

3. Kiss G, Van Cleynenbreugel J, Thomeer M, Suetens P, Marchal G.Computer-aided diagnosis in virtual colonography via combina-tion of surface normal and sphere fitting methods. Eur Radiol2002;12:77–81.

4. Paik DS, Beaulieu CF, Rubin GD, Acar B, Jeffrey RB, Yee J, Dey J,Napel S. Surface normal overlap: a computer-aided detectionalgorithm, with application to colonic polyps and lung nodules inhelical CT. IEEE Trans Med Imaging 2004;23:661–675.

5. Cathier P, Periaswamy S, Jerebko A, Dundar M, Liang J, FungG, Stoeckel J, Venkata T, Amara R, Krishnan A, Rao B, GuptaA, Vega E, Laks S, Megibow A, Macari M, Bogoni L. CAD forpolyp detection: an invaluable tool to meet the increasing needfor colon cancer screening. CARS 2004–Computer AssistedRadiology and Surgery, Proceedings of the 18th InternationalCongress and Exhibition, June 23–26, 2004, Chicago, IL,2004:978–982.

6. Kiraly AP, Laks S, Macari M, Geiger B, Bogoni L, Nova CL. Afast method for colon polyp detection in high-resolution CTdata screening. CARS 2004–Computer Assisted Radiologyand Surgery, Proceedings of the 18th International Congressand Exhibition, June 23–26, 2004, Chicago, IL, 2004:983–988.

7. Summers RM, Jerebko AK, Franaszek M, Malley JD, Johnson CD.Colonic polyps: complementary role of computer-aided detectionin CT colonography. Radiology 2002;225:391–399.

8. Pickhardt PJ, Choi JH. Electronic cleansing and stool tagging in

hip to a Haustral Fold, or Location in the Colon

Adenomas

�8 mm �10 mm

16/21 (76.2% [52.8–91.8]) 6/6 (100.0% [54.1–100.0])0/0 0/0

20/25 (80.0% [59.3–93.2]) 17/20 (85.0% [62.1–96.8])

16/22 (72.7% [49.8–89.3]) 11/13 (84.6% [54.6–98.1])26/30 (86.7% [69.3–96.2]) 14/15 (93.3% [68.1–99.8])

22/29 (75.9% [56.5–89.7]) 12/14 (85.7% [57.2–98.2])20/23 (87.0% [66.4–97.2]) 13/14 (92.9% [66.1–99.8])

expressed as number/total number (% [95% CI]) based on adenomas, and colonic location were determined at optical colonoscopy. Theitivities are not shown for these polyps according to shape, althoughocated on a haustral fold.verse colon, inclusive.

ions

t arel foldsensare ltrans

CT colonography: advantages and pitfalls with primary three-

1

2

2

2

2

2

2

2

2

2

2

3

3

3

3

3

3

3

3

3

3

4

4

4

4

4

4

4

4

4

4

5

5

5

5

5

December 2005 COMPUTER–AIDED POLYP DETECTION 1843

dimensional evaluation. AJR Am J Roentgenol 2003;181:799–805.

9. Pickhardt PJ, Nugent PA, Mysliwiec PA, Choi JR, Schindler WR.Location of adenomas missed by optical colonoscopy. Ann InternMed 2004;141:352–359.

0. Iordanescu G, Pickhardt PJ, Choi JR, Summers RM. Automatedseed placement for colon segmentation in computed tomographycolonography. Acad Radiol 2005;12:182–190.

1. Summers RM, Franaszek M, Miller MT, Pickhardt PJ, Choi JR,Schindler WR. Computer-aided detection of polyps on oral con-trast-enhanced CT colonography. Am J Roentgenol 2005;184:105–108.

2. Yao J, Summers RM. Three-dimensional colonic polyp segmen-tation using dynamic deformable surfaces. San Diego, CA: SPIEMedical Imaging, 2004:280–289.

3. Jerebko AK, Malley JD, Franaszek M, Summers RM. Computer-aided polyp detection in CT colonography using an ensemble ofsupport vector machines. CARS 2003–Computer Assisted Radi-ology and Surgery, Proceedings of the 17th International Con-gress and Exhibition, June 2003, London, England, Elsevier,2003:1019–1024.

4. Malley JD, Jerebko AK, Summers RM. Committee of supportvector machines for detection of colonic polyps from CT scans.San Diego, CA: SPIE Medical Imaging, 2003:570–578.

5. Gur D, Wagner RF, Chan HP. On the repeated use of databasesfor testing incremental improvement of computer-aided detectionschemes. Acad Radiol 2004;11:103–105.

6. Yao JH, Campbell S, Hara AK, Summers RM. Progressive featurevector selection scheme for computer-aided colonic polyp detec-tion. Presented at: RSNA Scientific Assembly and Annual Meet-ing, November 2004, Chicago, IL, 2004:633.

7. Bitter I, Brown JE, Brickman D, Summers RM. Large scale vali-dation of a computer aided polyp detection algorithm for CTcolonography using cluster computing. San Diego, CA: SPIE Med-ical Imaging, 2004:290–294.

8. Chakraborty DP. The FROC, AFROC and DROC variants of the ROCanalysis. Chapter 16. In: Beutel J, Kundel HL, Van Metter RL,eds. Handbook of medical imaging. Bellingham, WA: SPIE Press,2000:771–796.

9. Vining DJ, Shifrin RY, Grishaw EK, Liu K, Gelfand DW. Virtualcolonoscopy (abstr). Radiology 1994;193(P):446.

0. Yee J, Akerkar GA, Hung RK, Steinauer-Gebauer AM, Wall SD,McQuaid KR. Colorectal neoplasia: performance characteristicsof CT colonography for detection in 300 patients. Radiology2001;219:685–692.

1. Ferrucci J, Barish M, Choi R, Dachman A, Fenlon H, Glick S, LaghiA, Macari M, Morrin M, Paulson E, Pickhardt PJ, Soto J, Yee J,Zalis M. Virtual colonoscopy. JAMA 2004;292:431–432; authorreply 433.

2. Halligan S, Taylor S, Burling D. Virtual colonoscopy. JAMA 2004;292:432; author reply 433.

3. Pickhardt PJ. Virtual colonoscopy. JAMA 2004;292:431; authorreply 433.

4. Summers RM, Bitter I, Petrick N. Virtual colonoscopy. JAMA2004;292:432–433.

5. Macari M, Milano A, Lavelle M, Berman P, Megibow AJ. Compar-ison of time-efficient CT colonography with two- and three- dimen-sional colonic evaluation for detecting colorectal polyps. AJRAm J Roentgenol 2000;174:1543–1549.

6. McFarland EG, Brink JA, Pilgram TK, Heiken JP, Balfe DM, HirseljDA, Weinstock L, Littenberg B. Spiral CT colonography: readeragreement and diagnostic performance with two- and three-di-mensional image-display techniques. Radiology 2001;218:375–383.

7. Gluecker T, Meuwly JY, Pescatore P, Schnyder P, Delarive J,Jornod P, Meuli R, Dorta G. Effect of investigator experience in CT

8. Taylor SA, Halligan S, Burling D, Morley S, Bassett P, Atkin W,Bartram CI. CT colonography: effect of experience and training onreader performance. Eur Radiol 2004;14:1025–1033.

9. Pickhardt PJ. Three-dimensional endoluminal CT colonography(virtual colonoscopy): comparison of three commercially availablesystems. AJR Am J Roentgenol 2003;181:1599–1606.

0. Fletcher JG, Johnson CD, Welch TJ, MacCarty RL, Ahlquist DA,Reed JE, Harmsen WS, Wilson LA. Optimization of CT colonogra-phy technique: prospective trial in 180 patients. Radiology 2000;216:704–711.

1. Jiang Y, Nishikawa RM, Schmidt RA, Toledano AY, Doi K. Poten-tial of computer-aided diagnosis to reduce variability in radiolo-gists’ interpretations of mammograms depicting microcalcifica-tions. Radiology 2001;220:787–794.

2. Freer TW, Ulissey MJ. Screening mammography with computer-aided detection: prospective study of 12,860 patients in a com-munity breast center. Radiology 2001;220:781–786.

3. Paik DS, Beaulieu CF, Jeffrey RB, Yee J, Steinauer-Gebauer AM,Napel S. Computer aided detection of polyps in CT colonography:method and free-response ROC evaluation of performance (ab-str). Radiology 2000;217:704.

4. Gokturk SB, Tomasi C, Acar B, Beaulieu CF, Paik DS, Jeffrey RB,Yee J, Napel S. A statistical 3-D pattern processing method forcomputer-aided detection of polyps in CT colonography. IEEETrans Med Imaging 2001;20:1251–1260.

5. Acar B, Beaulieu CF, Gokturk SB, Tomasi C, Paik DS, Jeffrey RB,Yee J, Napel S. Edge displacement field-based classification forimproved detection of polyps in CT colonography. IEEE Trans MedImaging 2002;21:1461–1467.

6. Kiss G, Van Cleynenbreugel J, Thomeer M, Suetens P, Marchal G.Computer aided diagnosis for virtual colonography. Proceedingsof the 4th International Conference on Medical Image Computingand Computer-Assisted Intervention. Utrecht: The Netherlands,Springer-Verlag, 2001:621–628.

7. Yoshida H, Nappi J. Three-dimensional computer-aided diagnosisscheme for detection of colonic polyps. IEEE Trans Med Imaging2001;20:1261–1274.

8. Yoshida H, Nappi J, MacEneaney P, Rubin DT, Dachman AH.Computer-aided diagnosis scheme for the detection of polyps inCT colonography. Radiographics 2002;22:963–979.

9. Nappi JJ, Frimmel H, Dachman AH, Yoshida H. Computerizeddetection of colorectal masses in CT colonography based onfuzzy merging and wall-thickening analysis. Med Phys 2004;31:860–872.

0. Summers RM, Beaulieu CF, Pusanik LM, Malley JD, Jeffrey RB Jr,Glazer DI, Napel S. Automated polyp detector for CT colonogra-phy: feasibility study. Radiology 2000;216:284–290.

1. Summers RM, Jerebko A, Franaszek M, Malley JD, Johnson CD.Complementary role of computer-aided detection of colonic pol-yps with CT colonography. Radiology 2002;225:391–399.

2. Mani A, Napel S, Paik DS, Jeffrey RB, Yee J, Olcott EW, ProkeschR, Davila M, Schraedley-Desmond P, Beaulieu CF. Computedtomography colonography: feasibility of computer-aided polyp de-tection in a “first reader” paradigm. J Comput Assist Tomogr2004;28:318–326.

3. Summers RM, Dempsey J, Campbell SR, Yao J, Franaszek M,Brickman D, Dwyer A, Hara A. CT colonography with intravenouscontrast enhancement: computer-aided polyp and mass detec-tion (abstr). Am J Roentgenol 2004;182:75.

4. Yoshida H, Nappi JJ, Frimmel H, Miller FH, Dalal KA, DachmanAH. Computer-aided detection of polyps in CT colonography:performance evaluation based on combination of independentdatabases. Presented at: RSNA Scientific Assembly and An-

5

5

Di2

W

tVuL(

ScrsMR

1844 SUMMERS ET AL GASTROENTEROLOGY Vol. 129, No. 6

5. Bogoni L, Jerebko A, Dundar M, Lee J, Baker M, Macari M. Amultisite study to evaluate performance of CAD in polyp de-tection. Presented at: RSNA Scientific Assembly and AnnualMeeting, November 2004, Chicago, IL, 2004:577.

6. Franaszek M, Summers RM, Pickhardt PJ, Choi JR, Schindler W.Assessment of obscured colonic surface in CT colonography.Presented at: RSNA Scientific Assembly and Annual Meeting,November 2004, Chicago, IL, 2004:618.

Received June 15, 2005. Accepted August 17, 2005.Address requests for reprints to: Ronald M. Summers, MD, PhD,

iagnostic Radiology Department, National Institutes of Health, Build-ng 10, Room 1C660, 10 Center Drive MSC 1182, Bethesda, Maryland0892-1182. e-mail: [email protected]; fax: (301) 451-5721.

P.J.P.’s current affiliation is: Department of Radiology, University ofisconsin Medical School, Madison, Wisconsin.This research was supported by the Intramural Research Program of

he National Institutes of Health, Warren G. Magnuson Clinical Center.iatronix supplied the V3D Colon software free of charge. This studysed the high-performance computational capabilities of the Biowulfinux cluster at the National Institutes of Health in Bethesda, Marylandhttp://biowulf.nih.gov).

The authors thank William R. Schindler, DO (Naval Medical Centeran Diego, San Diego, CA) for providing computed tomographicolonography and supporting data; Andrew Dwyer, MD, for criticaleview of the manuscript; Shawn Albert and Tina R. Scott for databaseupport; Nicholas Petrick, PhD, for helpful discussions; Maruf Haider,D, and Meghan Miller for additional image analysis; and Sharonobertson for manuscript preparation.

Krukenberg of the Krukenberg Tumor

Friedrich Ernst Krukenberg (1871–1946) was born in Halle, Germany,into a family with a prominent medical lineage. His grandfather was theGerman anatomist Johann Christian Reil (1759–1813) for whom an areain the brain is named. Krukenberg began his studies in Halle, thentransferred to the medical school at Marburg where at the age of 24 hewrote a classical thesis on maligant tumors of the ovary. Thus began hislifeling interest in gynecologic pathology. In 1896, he described 5 cases ofwhat he took to be unique form of ovarian neoplasia, “. . .signet-ring cellsin a stroma of sarcoma.” Only later was this recognized as an anaplasticcarcinoma metastatic from the stomach. Despite Krukenberg’s misappre-hension, the eponym was perpetuated.

—Contributed by WILLIAM S. HAUBRICH, MDThe Scripps Clinic, La Jolla, California