Mass Detection in Digitized Mammograms Using ... - Semantic Scholar

2 downloads 0 Views 676KB Size Report
Computer-Assisted. Diagnosis. Schemes. Bin Zheng 1. Yuan-Hsiang. Chang. David Gur. OBJECTIVE. Using two independent computer-assisted diagnosis.
Mass Detection in Digitized Mammograms Using Two Independent Computer-Assisted Diagnosis Schemes

Bin Zheng 1 Yuan-Hsiang David Gur

OBJECTIVE.

Chang

Using

investigated

‘or’

the

operation MATERIALS

bandpass search

for

and

a single

suspicious

RESULTS.

in the

In this

(CAD-

per image. results

yielded

but

sensitivity

100%

CONCLUSION. upon the relevant

decreased Similar

10 years.

and

been

them

focused

diagnosis

15-81.

grams provide

cancer after revision

All authors: Department of Radiology, A461 Scaife Hall, University of Pittsburgh, 3550 Terrace St. Pittsburgh, PA 15261-0001. Address correspondence to B. Zheng.

0361-803X/96/1

to, morphologic

© American

AJR:167,

Roentgen

December

neural

analysis

trast Ray Society

1996

sion

[15].

networks

[4. adaptive

enhancement tree

than

schemes

clinical

useful

been

of

on

second the

of breast

191. Many implemented

CAD

rate of 0.79 of 94%

per

image.

to 0.4

including.

local

but

and

A logiper

[8].

[161.

and

binary

and

image.

depending

by combining

that

required

To

achieve

use.

more

schemes

produce

tion

(typically,

rate

clusters regions

[ I 7. 18]. How-

analysis performance

of

high

than a high

to small

CAD

image

detec-

per

image

for

and two These results

databases.

Multiple

independent readings of the same films several readers (radiologists) can improve diagnostic

improvement

detection

and

sensitivity [19.

20]. in

specificity

However.

the

performance

results

of several

by

con-

Two

CAD

deci-

detection

not to

been

independent in

developed

digitized

by the

of the potential combining

CAD

investigated

multilayer

sensi-

most

false-positive

I .5 regions

is

accepted

detection

85%).

of microcalcifications per image for masses).

are limited

CAD

fbr widely

been adequately our knowledge.

edge-gradient

density-weighted

approach

be improved

overall

(i.e..

tivity

system

filtering [ ID]. arti[ 1 1-1 3]. wavelet-trans-

filtering analysis

a

the combined

rate

feature

lower

to improve

schemes,

141.

of 2.07

can

in the last

mammoeventually

environment

in various

ficial

a

have

not limited

analysis

operation.

double-reading

or specificity

the

Some

of diagnosis

and tested

rate

428 and

compared.

a sensitivity

detection

ever,

been

concentrated

tool

accuracy

techniques

analysis

676-1421

with

in a screening

advanced

form AJR 1996;167:1421-1424

and

detection

were

detection

an “or”

false-positive

included schemes

feature

achieved

With

topographic

(CAD)

of microcalcifica-

or a prescreening

efficiency

Received February 23, 1996; accepted June 13, 1996.

tested.

masses in digitized CAD schemes could

radiologists

opinion

CAD

and

[ 1-4]. and others of

operations

topographic

(CAD-2)

two

Gaussian

a five-stage

schemes.

has

different

on detection

clusters

detection

“and’S

with

that

of the

we

a logical

with

database

and had a false-positive

independent

CAD

extensively

many

of the

the other

image

or logical

a false-positive

sensitivity

mammography

developed

“or”

scheme

an

independent

oniputer-assisted

have

tion

to

question.

investigated

schemes.

applying

to 90%.

clinical

of digital

by

performance

of I .69 per image.

with

and

to a large

a multilayer

a reduction

(CAD)

detection

analysis

The

of 96%

method rate

sensitivity

of several

C

study.

search

produced

also

applied

logical

a sensitivity

mass

using a logical “and” operation. independent mass detectors. one

masses.

ofeither

detection

operation

results

form

diagnosis

of

feature

were verified

preliminary

A five-stage

“and”

sensitivity

topographic

220

I ) achieved

a false-positive

cal

computer-assisted the

region.

with

ofthem

scheme

the

multilayer

mammograms

combination

had

to improve

and to improve the specificity AND METHODS. Two

filtering

digitized

independent

two

potential

schemes

has

in this

area

schemes

for mass

mammograms

have

in our laboratory

118. 21). The

1421

Zheng

two

schemes

use

independent

approaches

select the initial suspicious regions and analyze image features in those regions. first

scheme

detects

suspicious

regions

to

90

then The

80

based

bandpass filtering and multilayer feature analysis [ 18], and the sec-

on Gaussian topographic ond scheme

applies

to search

for

five filtering-based

a specific

type

stages

at each

stage

using

a large

image

results

from

the

independent

schemes

the effect

of a logical

compared

“or”

and a logical

Materials

to assess “and”

60

In this

mu

study,

two

70

of suspicious

region

were

[21].

preliminary

database,

the 40

E

combination.

z

30

and Methods

20

Four hundred twenty-eight images were selected and used in this study. All images were acquired during

the

breast

examinations

of

178

patients

10

at the

University of Pittsburgh Medical Center and its affiliate clinics and hospitals. With the exception of 34 examinations, all examinations (each of which was positive for a mass) were selected from a biopsy listing.

To stress

the system,

one radiologist

the 34 true-negative examinations showed asymmetric tissue structures glandular negative

tissues. cases

The

inclusion

enabled

5

10

15

of these

20

Effective

selected

because they or dense fibro-

us to evaluate

0 30

25

35

40

50

45

Size of Mass (mm)

Fig. 1.-Bar graph shows distribution of effective size of true-positive square root of product dimension of mass (maximum times minimum).

masses. As defined here, effective size is Total database includes 220 mass regions.

difficult, the

perfor-

of the CAD schemes in various true-negative cases, including those considered subtle or difficult mance

by experienced

et aI.

lOO

observers.

For 194 of the selected images. a contralateral image of the same view (i.e., left and right craniocaudal or mediolateral oblique view) was included in the

study.

The

remaining

40

images

included

only a unilateral mammogram that was available for various clinical reasons in these specific examinations. In summary, 220 mass regions were identified and visible in this set of 428 images; 1 15 of these regions showed different views of 82 malignant masses and 105 showed different views of 75 benign masses. Of these 220 masses, 77 were well circumscribed, 82 were partially circumscribed, and 61 were spiculated. Tissue patterns surrounding the masses varied. Predominantly

fatty,

heterogeneous,

and

a)

50

2 z

dense

I

fibroglandular tissue patterns surrounded 84. 87, and 49 mass regions, respectively. The distributions of mass characteristics in this image database, including the size and digital value contrast of all true-positive masses, can be found elsewhere (22] and are summarized in Figures 1 and 2. The mass locations in the images were identified and marked by mammography experts. Mammographic

MinR

were

acquired

using

gray-level

and MRE-l

NY)

combination

resolution.

After

digitization.

200 Mass

Fig. 2.-Bar in digital

400

Contrast

600 (Digital

800 Value

1000

1200

Difference)

graph shows values between

itself. (Dimension

distribution of contrast oftrue-positive masses. As defined here, contrast is difference mass region and square window surrounding mass region, but excluding mass region of background window was 20 pixels [8 mm) larger than maximum axis of mass region.)

the

images were subsampled by a factor of four in both dimensions to make the size of digitized mammograms approximately 600 x 450 pixels. A

1422

0

a

film (Eastman KOdak, and were digitized using a laser film digitizer (Lumisys. Sunnyvale, CA) with a pixel size of 100 x 100 .tm and 12-bit Rochester,

screen

films

#{163}

detailed description of the two CAD schemes in this study, including the computational rithms,

optimization

procedures.

and

preliminary

used algo-

testing

results,

has

been

reported

[18,

21].

How-

ever, the image database used in this study is different from that used earlier. Among the 220

AJR:167,

December

1996

Mass verified

mass

cases.

not used

were

Each tion

by the two

results

I I 8. 2 I 1.

CAD

were

coIputer

than

in the database

niage

pendentlv

(more studies

I-tf4

previous

iii

was

in Digitized

two

thirds)

processed

mdc-

regions

false-positive

image.

detection

The

study.

which

were

were

different.

operation.

cmous regmomis

the combined

a l(XLI

point

niinimum

region-gross

ing

Because

growth

region

regmons).

ferent the two tamice

schenics

schemiies

the

distance

and

axis

third

than

regmt)n.

consmdered

tO

the

suspicious tVi)

for

miiass

or lx)th:

scliemiies

a region

remained

detected

b

If this of

were

masses, were

885).

The

logical

( I l/

of the mass

However.

the

false-positive

total

of885).

which

CAD

scheme

tering

and

using

Gaussian

multilayer

[ I 8].

analysis scheme

using

[2 1 ].

Using

rately,

we

with

the

and

1 .69

logical ity

CAD-I

was

a five-stage

image

quality

renimmmied

operatmon. if it was

the

two

achieved

“or”

image.

94%

are

not

large

the

and

diverse,

the to

lOOC/c

detection with

feature

that

it)’ and

does

consistently

not generate

a large

same

study

[22].

the

classification

feature

were

designed

the

difference

to CAD-

at i false-positive

false-

image

22) versus

positive

detection

developed

several

96%

and

detection

rate

sensitivity [this

of 0.85

at 0.79

per

schemes by different

been

research

and

would

increase (25).

rate

in CAD-2

( I .69 I 0.79

2. 14)

=

differences schemes

resulted

positive

detection

operation

was

pose.

the

with

pair

of

“or”

and

solution could

were

for actual

same

an

of

detected

339

l’alse-positive

regions

December

1996

reduce

the

maintaining

Although image

regions

greater than

false-positive

regions

CAD-2) by

findings

mdi-

in feature

or variability a

Hence.

a high

I 73

by

At false-

identified

These

effect

on

the

the true-posi-

on

combining

independent

results

schemes detection

of

could rate

while

sensitivity.

this database density

were

CAD-I

noise

more

2 1 1 truewhich

overlap).

had

or

scale

(93%

characteristics

detections.

if scheme ordinal

findings.

197

overlap).

image

[19]. This

I detected

the

or may

by CAD-2

with

two

with the probability

or negative

time.

that

than

operations)

using

regions.

(51%

both.

scheme

weighted

facilitated

CAD-

detected

not

more

(i.e..

that correlates

mass

also

but

of

better

positive

CAD-2

AJR:167,

pur-

this

for this purpose

be

Interestingly.

in

“or”

characteris-

of each

“and”

provided

overlapped

istedi diagnosis, CAD-i = CAD scheme using a multilayer topographic feature analysis, CAD-2 = CAD scheme usmnga five.stage search method, CAD-i or CAD-2 = region is identified as a mass by either CAD-i or CAO-2 or both, CAD-i and CAD2 = region is identified as mass by both CAD-i and CAO-2.

false-

for

positive regions. of which only (out of 719 regions detected

e

two

independent

appropriate

or specificity. of the performance

(not binary)

two

I).

the

the logical

sensitivity

results

tive

twice

combined

when

I 19. 25). Logical operation independent CAD schemes

cate

than I (Table

perfbmmance

more

false-

between

A

similar be

and

the

because a logical operation based on two independent schemes could improve

either

the

more

of CAD-

a large

used.

“and”

specificity

was

rate

and

logical

Because

that

in

of

sensitivity

perfoniiance

in

would

use

that

sensitivity

positive

have

CAD

a logical

decrease

false-

study).)

a means of

the

operation

approach

was

provide

increase

specificity

be a better

in the two

difference

or

sensitiv-

that

nonweighted

a nonadap-

sensitivity

CAD

independently

conditions

detection

per image

as

in the

of two

to analyze

to increase

It was anticipated “or” operation would

regardless

of performance

an adaptive

(The

analysis

boundary

A slight

resulted.

criteria

because.

for a comparison between

Because

different

1 2 specific

tive optimization.

sensitiv-

of

use

diagnos-

need

schemes

was used

database

the

on the relevant

the

tics

number

iniprove

schemes

overall (i.e..

the same

(to

ity or specificity).

only image

study.

the

a CAD sensitiv-

detections. the

other

93C/(

where

l0(Y/

that

CAD

depending

positive

Consis-

to pursue

suggests

improving

tic question a

characteristics

yields

fil-

of 0.79

2.07

tissue

topographic

sensitivity

study

of

I 19). this pre-

perl’oniiance)

independent

techniques

this frac-

environment,

it is impractical

scheme

studies

diagnostic

to that

to the use

radiologists

performance

of cases

to ensure

enough

clinical

Similar readings

might

alone yielded two previous

overlapped.

I was

sepa-

Using

studies

schemes.

of

reduction.

a fraction

related

method

rate

respectively.

operation.

increased

and detection

an 8l/

of the multilayer

feature

schemes

was

by

seeni with

the samie set of images

elimi-

overall

and

in

(froni

prelimithe

comparable

more

database).

operation

manner

to

missed

detections

the

bandpass indicates

CAD

96%

false-positive per

experimental

searching

in this

I 18. 2 1 1. Although

In

represents

CAD-2

were

a

for other

liminary

detec719

resulted

that

results

was

topographic

and

overall

reports

tency.

by emther of the

the

study.

all of the 220

that either CAD scheme that are consistent with

Note

tion

schemes.

summarizes

this

several

Discussion

in another

1

inmges

in

decrease

corn-

Results

in

operation.

“and”

712

‘and

only

independent

operation

logical

nated

thcmt

a regmon

of multiple

(from

24)).

independently

sen-

23%

regions

developed

“or”

logical

false-positive

by

were

scheme.

“or”

niasses

in the three

were

the

“and”

true-positive

Although

Table

but

addressed.

per

shown

have [23.

of these CAD techniques must be The schemes used in this study

performance

small

which

usefulness

by the 100%

included

increased

of

clinical

the

either

logical

results

ftlse-positive

results

the

used

or a Iogmcal

fir mass

verified

23

through

Using

region in our

(some

nary

perform reported

by

as indicated

The detec-

in the logmcal “and

CAD

axis

regIons

detected

suspicious

both

I Il.

om. overlappmng.

1 it was

dms-

as corn-

5U5tCtOU5

resUlts tromi#{236} the t5’’() scliciiies bined usmng emther a logical ‘or’ operatmon. Iii the “or’ operatmon.

dms-

regIon

layer

tion

to 0.4

Diagnosis

groups

logical

LX)th

the

half of the longest

tWO

the sanie

tiC

the by

wmth

the grosvtli

ni

dif-

whetliem

regmomis.

compared

false-

mdentmfy

mdemitmficd

topographic

was snialler

the growth

same

seeds

was calculated by

filters

To determiimne

the

two

tance of the longest puted

seeds.

the

miiay

schcmiies

mdentmfv

between

difThremit

for tn.me-pt)smtmve and

as growth

poInts

each

the

niasses missed

attained

sitivity

tions

in a data inside

somitewhat

two

the

for

coordinates

fluctuation

amid the

for

seed

the

seed are recorded

thresholding

before

posItIve

searches

growth

and

of the density

suspicious

used

as an initial

algonthms.

)Iocatmon I of ever’, file.

schenie

Computer-Assisted

using

rate

true-positive

each intage for regions 5U5ICtOU5 for a mass that were identified by bch schemes. Although the two schenies use difierent approaches. including dm1ferent region growth algorithms to identif) suspithe mmmgc. each

image:

ttmtonatmcally

to

search

Iii

per

Using

operation. the detection sensitivity reduced to 90%. with a reduction in the

was

and the detecindependently. A

siittefl

W15

Mammograms

positive

“and”

schemes.

recorded

programli

Detection

patterns

is large and

and mass

diverse feature

1423

Zheng

characteristics,

selected

note

small,

the and

biased

potentially

masses

in the tested

in the actual results.

Therefore,

study

with

from

blinded

8.

is required

Radiolog 9. Vyborny

findings.

I. Fam BW, Olson

SL. Winter SF. Scholz FJ. Algorithm for the detection of fine clustered calcifications on film 169:333-337

2. Chan

HP,

mammograms. Vyborny

detection

Ci.

grams work.

1988:

microcalci-

1385-1390 4. Qian W, Kallergi M, Clarke LP. et al.Tree

neural

struc-

classificationof breast tumors. Comput Doi

K, Vybomy

of bilateral-subtraction

CJ.

and

Lu

P. Huo

extrac-

neural

pattem

diagnosis itive

B, Lam

MedPhvs

1995; MA,

breast

20.

21.

with

and sin-

enhancement

Acad

Computerized mammograms

Acad

a multilayer

1995:

Radio!

adaptive filtering.

mammography

KL,

Helvie

MA.

22:1555-1567

on digital

main-

RM. Jiang and

Y. Papaioan-

24.

computer-aided

reduction

local

mograms.

Acad

Nishikawa

RM,

masses

on

digital

density-weighted Proc

mt

mammocontrast

Soc

Opt

Eng

Radio!

AS.

Benefit

in a populationprogram.

for

masses

Radi-

identiin digi-

1996:31:143-

Radio!

clinical

“intelligent”

Nishikawa

RM, DE,

experience

Papaioannou

with

mammography

diagnosis. Haldemann Schmidt

J.

a prototype workstation

Proc

mt Soc

RC.

Giger

detection

microcalcifications

CA, double

Radiolog’

1995:

Sullivan reading

K.

Opt

ML,

Perfor-

scheme

on a clinical

for mam-

for computer-aided

workstation

Radiology

Doi

RA,

of a computerized

Beam the

14

RC,

1995:2434:65-71

Wolverton

nosis.

1996:3:8()6-8

Heldemann

ML. Initial

mography 25.

Med

Taube

regions Insect

Giger

clustered

edge-

analysis.

Zheng B. Chang YH, Our D. An adaptive computer-aided diagnosis scheme of digitized mam-

mance

of false-pos-

using

of suspicious mammograms.

Emg

analysis

KA. reading screening

for computer-aided

Med

analysis.

of ROC

rep-

prediction

Chang YH, Zheng B. Our D. Computerized

150

23.

from

images:

I 2:6(1-75

double

1994:191:241-244

tized

micro-

in accuracy

EL, Lemevall

ologv

Eng

an artificial

Gains

1992:

based

enhanced

et al. Classification

microcalcifications

using

Thurfjell

Opt

texture

Nishikawa

of breast

grams

Our D.

microcal-

and

analysis.

in terms

Making

fication

gradient analysis.Med Pins 1995:22: l6l-l69 16. Petrick N, Chan HP, Sahiner B. et al. Automated

Schmidt

D.

ofdiagnostic

of independent

13

feature

readings

Decis

Soc

with

tissue

in mammography:

clustered

Our

segmentation

feature

W.

mammograms.

in digitized

single-image

heated

perfor-

of mammographic recognition

HP, Helvie

Doi K,

T,

J. Image

masses

19. Metz CE. Shen JH.

net-

Z. Computer-

lot

multiresolution

nou

of

topographic

Pme

lesions:

network

Sahiner

Pin’s 1995:22:1501-15 Ema

M. Good of clustered

digitized

and assessment

artificial

Proc

and normal

mograms:

using

22.

network.

detection

1992;25:218-237

Yin FF, Giger ML, RA. Comparison

1424

Biomed

ML.

detection

D, Chan

of mass

15.

P.An approach

Wei

in

18. Zheng

AiR

Staiger

detection

1995:2:655-662 B, Chang YH.

Radial

I905: 702-715

neural

HP, La SB,

calcifications: 14.

to automated detection of tumors in mammograms. IEEE Trans Med Imaging 1990;9:233-24l 6. Ng SL, Bischof WF. Automated detection and Res

extraction.

Computer-aided

wavelet transform segmentation of microcalcifications in digital mammography. Med Pin’s Brzakovic

vision

mammography.

of mammographic

of artificial

YH.

2:959-966

M, Giger

feature

B. Chang

cifications

Hil-

1995:2434:598-6tJS

tured

1995;22: 1247-1254 5. Brzakovic D. Luo XM.

PD.

Computer

in

1993:

mance

13. Chan

3. Davies DH, Dance DR. The automatic computer detection of subtle calcifications in radiographically dense breasts. Pits’s Med Biol 1992;37:

7.

Kupinski

Zheng

detection

using a shift-invariant MedP/mvs 1994:21:517-524

ized detection

fications on mammograms: the potential of computer-aided diagnosis. Invest Radio! 1990;25: 1102-1110

Bourland

ML. Computer-aided for spiculated lesions.

I I. Zhang W, Doi K. Giger ML, Wu Y. Nishikawa RM. Schmidt RA. Computerized detection of clustered microcalcifications in digital mammo-

et al. Improvement

of clustered

JM.

10. Thao D. Rule-based morphological feature tion of microcaicifications in mammograms.

12.

Doi K,

in radiologists’

Radiology

Pruneda

MW. Nipper screening

Opt Etig

1995:2434:590-597 I 7.

Coniputer-aided

1994:191:331-337 CJ, Giger ML.

!imt Soc

References

I,mre.st

masses.

I

artificial intelligence 1994:162:699-708

posi-

to validate

WP.

lis A. Riggs mammographic

that the

Kegelmeyer

in the computer-

techniques

of mammographic

1993:28:473-48

Radial

prospective

of both

processing

ized detection

a

preva-

may affect

follow-up cases

that

differs

a large

tive and negative

fact

population

adequate

gle-image

was but

of detectable

The

database

imaged

database represents

subset

on mammograms.

lence

these

that

retrospectively

et al.

diag-

l97(P):425 DC.

What

arc the issues in

of mammograms’?

(letter).

1994:193:582

AJR:167,

December

1996

Suggest Documents