Heber MacMahon, MD #{149}Kunio Doi, PhD Maryellen L. Giger, PhD #{149}Charles E. Metz, Xin-Wei Xu, BSc #{149}Hisashi Yonekawa, MSc4

Data

Compression:

Accuracy

D

compression

ratios

in randomized gists.

The

were

order results

each idly

Radiography’

hundreds

of such

working day, and inexpensively

the

images

ability transmit

RE0ONSTRUCRON

in

to rapand

tune archiving and communications systems. Therefore, extensive mesearch has been performed in recent

years to develop data compression techniques that allow a digital image to be represented by a reduced amount of digital data (2-4). Howevem, significant image deterioration can occur at high compression ratios. Therefore, for a given compression scheme, it is important to determine the highest compression ratio that can yield acceptable clinical images. We have evaluated a new irreversible compression technique to determine

to 12 radiolothat,

coM’RESsCN

IGITAL

with

its effect on diagnostic chest radiology.

MATERIALS The

data

based coding

Index terms: Data compression #{149} Images. processing #{149} Picture archiving and communication system (PACS) #{149} Radiography. digital, 60.1215. Radiography, technology #{149} Receiver operating characteristic curve (ROC) 1991;

Chest

stone such large data sets has become an important challenge for designers of digital imaging systems and pic-

this compression scheme, compression ratios as high as 25:1 may be acceptable for primary diagnosis in chest radiology.

Radiology

on Diagnostic

generate

MSc

#{149}

chest radiography mequires high-resolution images in the mange of 2,048 X 2,048 pixels with 1,024 or more gray levels (1). Because large radiology departments

presented

suggest

PhD

Effect

in Digital

High-resolution digital images make up very large data sets that are relatively slow to transmit and expensive to store. Data compression techniques are being developed to address this problem, but significant image deterioration can occur at high compression ratios. In this study, the authors evaluated a form of adaptive block cosine transform coding, a new compression technique that allows considerable cornpression of digital radiographs with minimal degradation of image quality. To determine the effect of data compression on diagnostic accuracy, observer tests were performed with 60 digitized chest radiographs (2,048 x 2,048 matrix, 1,024 shades of gray) containing subtle examples of pneumothorax, interstitial infiltrate, nodules, and bone lesions. Radiographs with no compression, with 25:1 compression, and with 50:1

Sanada, BA2 #{149}Steven M. Montner, MD Nobuyuki Nakamoni, BSc3 #{149}Fang-Fang Yin, #{149}Hiroshi Takeuchi, PhD4

#{149} Shigeru

accuracy

AND

compression

on adaptive (4), which

in

block is an

is

cosine irreversible

transform tech-

Figure

1.

Basic

DATA

scheme

of compression

reconstruction of data from by means of adaptive block method. 2D = two-dimensional, verse OCT.

digital cosine

and

images transform

JDCT

in-

tion intervals of transform

and the coefficients

cutoff frequencies are determined

individually

for

image

each

block.

These

quency

distribution

cients,

respectively,

of transform as well

coeffi-

as the

com-

nique (Fig 1). A digital chest image (2,048 x 2,048, 10 bits per pixel) is divided into 16 X 16 block images. Transform coeffi-

pression ratio desired. At a given compression ratio, the optimal image quality can be obtained by appropriate selection of these two parameters. For example,

cients

small

are

obtained

by

means

of a two-di-

mensionab discrete cosine transform (DCT) (5). One of the unique features

178:175-179

IMAGE

parameters are estimated by taking into account the histogram and the spatial fre-

METHODS technique

CO4PRESSED

distinguish

this

proprietary

OCT

that

tech-

nique (Konica Corp. Tokyo) (6) from othem OCT methods (7) is that the quantiza-

quantization

intervals

and

off frequencies block images,

are employed whereas large

intervals

and

high

employed Selection

for rough of quantization

cutoff block

bow

cut-

for smooth quantization

frequencies images intervals

are (Fig and

2).

cutoff frequencies in this way tends to meduce artifacts in both smooth and rough I

From

The

Kurt

Rossmann

Laboratories

for Radiologic

Image

Research,

Department

of Radiolo-

gy, the University of Chicago, 5841 5 Maryland Ave. Chicago, IL 60637. From the 1989 RSNA entific assembly. Received June 5, 1990; revision requested July 25; revision received August accepted August 24. Supported by U.S. Public Health Service grant CA24806 and the Whitaker Foundation. Address reprint requests to H.M. 2

Current

address:

College

3

Current Current RSNA,

address: address: 1991

Kyoto Institute Research and

4

of Biomedical

Technology,

of Technology, Development

Kanazawa

University,

Kyoto, Japan. Center, Konica Corporation,

Kanazawa,

Tokyo,

sci20;

block high

images, compression

cant degradation To measure

thus

the

facilitating ratios without

use

of very signifi-

in image quality effects of various

(6). bevels

Japan.

Japan.

Abbreviations: form, ROC

=

DCT receiver

discrete operating

cosine transcharacteristic.

175

Cl, I-

z

w C-) LI. U-

w

0 C) U0

w D

z

I

I

:

TRANSFORM

: COEFFICIENT

-w-----i

k

WL QUANTIZATION

SPATIAL

INTERVAL

a.

FREQUENCY,

u

b.

Figure 2. (a) Histogram of transform coefficients for smooth and rough block images (due to small and large density variations, respectiveby) and the corresponding quantization intervals. Rough block images can be quantized with a barge quantization interval (WL) and thus a small number of quantization levels. Smooth block images can be quantized with a small quantization interval (W5). The higher the compmession ratio, the larger the quantization interval in general. (b) Spatial frequency distributions of transform coefficients for smooth and rough block images and the corresponding cutoff frequencies for removal of high-frequency components. Smooth block images contain very low-frequency components and thus can be represented by a small number of transform coefficients. In general, the higher the corn-

pression

ratio,

the lower

the cutoff

frequencies.

of data compression on diagnostic accuracy in chest radiography, we employed observer performance tests with receiver operating characteristic (ROC) analysis. For practical reasons, the format observer test was limited to a study of two compmession ratios. Initially, two experienced observers reviewed a chest radiograph containing multiple abnormal findings displayed with various compression ma-

scored separately. Thus, 60 madiogmaphs constituted 120 separate images for scoming purposes. Among these images were 20 hemithoraces with noncalcified pubmonary nodules, 20 with pneumothomax, 20 with interstitial infiltrate, and 20 with bone lesions. All abnormalities were visually subtle. For the purposes of the observer test, bone lesions were defined as either recent fractures on tumor involve-

tios

alongside

ment.

(Fig

3). Compression

the

uncompressed

image

more

than

viewed

between

was

the

same

were

the

unlikely. other

hand,

Oeteniomation was

judged

at to be

sufficiently marked that statistically significant bosses in accuracy were expected with use of our standard observer testing procedures. A total of 60 postenoanteniom and anteropostenior screen-film chest radiographs (Lanex Medium/OC system; Eastman Kodak, Rochester, NY) were digitized with a 2,048 X 2,048 matrix and 1,024 gray levels with use of a high-quabity drum scanner (Fuji Photo Film, Tokyo) (8),

and data

compression

was applied

with proprietary software (Konica). Each radiograph was then reconstructed and printed with a baser printer (Konica) at 2,048 X 2,048 X 10 bits with no compmession

and

50:1

ratios

after (Fig

compression 4). These

im-

ages were 12% smaller than the original radiogmaphs. Oensity and contrast of the compressed images were closely matched to those of the uncompressed originals. In the observer test, each side (ie, hemithomax) of each chest radiograph was 176

Radiology

#{149}

with

certain nate cific

additional

tomogmaphic scans follow-up examinations

verification.

As

images

in a previous

were

for the presence abnormalities

deemed

for (9),

of one or more spedue to a lack of conclu-

ative each

were different For pulmonary

hemithoraces) abnormality. were

such

find-

indetermi-

As a result, hemithoraces

ules, theme true-negative tial infiltrate,

data

and

study

sive proof of disease. number of control

20 true-positive hemithomaces; 20 true-positive

the (ie,

for nod-

and 71 for interstiand 63 true-

radiologists

bers, seven iting fellow) Three types compressed, pressed

senior residents, participated as of images were compressed at

at 50:1.

madiogmaphs ized viewer

(four

were with

For

the

faculty

and one visobservers. tested: un25:1, and com-

observer

displayed the view

mem-

test,

the

on a motorboxes masked

systematically out

possible

Eight

training

cases

beginning

of the

to familiarize scoring system

the and

first

radiolotest for-

mat. A maximum of 60 seconds was altowed for viewing each complete chest image. For each hemithomax, the presence or absence of each of the four findings was mated on a scale of 1 to 5, with 1 mdithe

definite

absence

and

5 the

defi-

nite presence of the abnormality in question. The observers’ verbal responses were recorded by another individual who monitored

the

observer

testing

scnibed

tests.

each

Other

details

procedure

previously

For neg-

negative; for pneumothomax, 20 true-positive and 100 true-negative; and for bone lesions, 20 true-positive and 95 true-negafive.

Twelve

at 25: 1 and laser-printed

together

as computed ings from

radio-

of three sesthe images

to cancel

at the

cating

seemed on

varied

session.

shown

graphs

50:1,

evalua-

Sixty

in each in which

learning effects (8). To further reduce learning effects, at least 1 week elapsed between each reading session, and the same chest was never shown twice during

curacy

objective

bight.

observers

tion based on this subjective review. These ratios were chosen because the degree of image quality loss at 25:1 cornpmession was judged to be so slight that significant degradation of diagnostic ac-

for

and

had

were

session, mainly gists with the

selected

of 25:1

hemithomaces

ambient

were viewed and the order

one type of abnormality, and 25 hemithomaces were completely normal. The presence on absence of each abnormality was established by consensus of two expenienced chest radiologists (H.M., 5MM.) who did not participate in the observer test, using the original radio-

50: 1 were

ratios

Some

to exclude graphs sions,

of the

have

been

de-

(9).

detection

task

and

each

type

of

image, a binommab ROC curve (10,11) was fitted to each observer’s confidence mating data by maximum likelihood estimation. The index A, which represents the area under plotted

a binommal in the unit

ROC curve when it is square (11,12), was

then calculated for each fitted curve. The statistical significance of apparent differences between ROC curves was determined

by

paired

data

applying

a one-tailed

to the

t test

reader-specific

A2

ues (8). Our use of a one-tailed based on the assumption that differences were present, the pressed images would not be cally and not

superior the 50:1 be

superior

test was if any real data-corndiagnosti-

to uncompressed compressed images to the

25:1

for

vab-

images would

compressed

January

1991

b.

C.

e. 3. (a-f) ing pneumothorax,

Oigitized lung

Figure

images.

Note

which

leaves

that

use of a two-tailed

open

test,

the possibility

that ei-

them image may prove superior, would have doubled the P values reported next. To represent the overall detection accumacy of the entire group of observers with each type of image in each task, composite

ROC

ing

the

curves slope

cept

parameters

cific

curves.

were

calculated

parameters

of the

and

by the

various

avemaginter-

reader-spe-

Accuracy

of pneumothorax

was

highest

with

.01)

with

the

50:1

detec-

among

observers,

pressed

images,

7). The

ROC

loss

in accuracy

Volume

178

was

Number

#{149}

in the

However, significant

1

differ-

the (P

with loss

50:1

for

showed

compressed (P = .03).

bone

images

and

25:1 or 50:1 8). Composite

compression ROC curves

data from interstitial detection

those ratios (Fig

=

images

with 25:1 difference 25:1 and

was

signifi-

DISCUSSION

lesion

between

compressed

significant

accuracy .36). The between

com-

no statistically

differences

cant

approached

.06). For interstitial none of the difthe ROC curves apsignificance (Fig

results

also

which combine thorax, nodule, and bone lesion

variation

but the

considerable

observers.

no statistically

50:1

significant

across

dicated

loss of accuracy with 25:1 compressed images fell short of statistical significance (P = .17). again due to variation

detection

uncompressed

images.

loss of diagnostic compression (P in overall accuracy

images (Fig 5). The accuracy with 25:1 compressed images was slightly lower, but this difference was not statistically significant (P .19), due to ence

compressed

victim with multiple abnormalities includbevels of data compression, as indicated.

In the case of nodule detection, accuracy was again highest with uncompressed images (Fig 6). The apparent

significance (P infiltrate detection, femences between proached statistical

RESULTS

tion

f.

chest radiograph (2,048 X 2,048 matrix, 1,024 shades of gray) of gunshot contusion, and bullet fragments. The same image is shown with increasing

un-

with (Fig 9),

pneumoinfiltrate, results, in-

The fact that there was a more pronounced loss of diagnostic accuracy for pneumothomax than for the other abnormalities in the compressed images is consistent with the loss of high-frequency components caused by

this

type

of compression

(Fig 10). The lack fect of compression cumacy for interstitial bone lesion detection lated to the specific

scheme

of a significant on diagnostic

efac-

infiltrates or is probably menature of the

Radiology

177

#{149}

b.

a. Figure The

4.

(a-c)

same

image

Figures

5, 6.

relative pulmonary

accuracy nodule

compressed

Digitized

chest

is shown

with

ROC

and

curves

radiograph 25:1

and

indicating

of pneumothomax (6) detection

25:1 and

(2,048 50:1

studied,

as the

septal

lines,

detection

such

in some

test

did

not

detection bone lesions

depend

on

fine

Their

ROC

pression

ratios,

and

the

I-

U-

U-

w

w >

0

0

0.

w

w

I-

0 0

0 FALSE

of many in this detail

results

mdi-

authors

con-

POSITIVE

FRACTION

are

progessive,

showing

no evidence

of an abrupt deterioration 20:1 and 30:1 ratios (Fig

Our

results

suggest

of this adaptive block form coding technique,

comparison nadiognaphs.

between

with

cosine data

use

transcom-

pression ratios as high as 25:1 may be acceptable for digital chest radiology. Although theme is a mild trend toward lower accuracy with 25:1 compressed images compared with uncompressed images, this trend was

cunning 178

losses

at high

Radiology

#{149}

in

image

compression

in this quality

ratios

oc-

1.0

sufficiently

across

marked

observers

or consistent

to approach

1.

statisti-

cal significance, even with the 12 observers and relatively large numbers of very subtle cases used in this study.

This

almost

ceptabbe

suggests

certainly

for long-term

with

subsequent

chest

#{149}

References

3.

the

0.8

FRACTION

Acknowledgments: We are grateful to the following radiologists who participated in the observer tests: Steven M. Montner, MD, Michael Jokich, MD, Carl J. Vyborny, MD, PhD, Charlene Sennett, MD, Mark Ridlen, MD, Tony Kim, MD, Brian C. Randall, MD, Chihiro Abe, MD, PhD, Mark Backus, MD, Bruce Goethe, MD, Donatus Siliunas, MD, and John Hegarty. MD.

and

et al (6) employed

study,

0.6

3).

that,

2.

ekawa

0.4 POSITIVE

6.

ages with 25:1 data compression may be acceptable for primary diagnosis,

of Yon-

0.2 FALSE

5.

res-

in image 25:1 com-

scheme

a).

C,,

study.

OCT

(arrow,

0

not

the

pneumothomax

0

cluded that 25:1 compressed images were unacceptable for clinical use. We believe that the differences between their results and ours are directly attributable to details of the compression schemes used in each With

a small

z

cases

In a recent study, Ishigaki et al (7) evaluated another proprietary OCT scheme, with significantly different cated a sharp deterioration quality between 20:1 and

with

as indicated.

I0

as

olution.

conclusions.

of a patient

0

depends on recognition of subtle large-area density differences in the lungs, which are relatively webl-pmeserved by this data compression

scheme. Similarly, of the metastatic

of gray)

> I(1)

preserva-

information

shades

0

tion by the data compression scheme of clinically relevant frequency components in these images. Though many interstitial infiltrates contain high-frequency

ratios,

1,024

z

50:1 compressed

as well

compression

matrix,

the

(5) and with un-

images.

cases

c. X 2,048

that

use

would

of im-

be ac-

archiving

for

MacMahon H, Vyborny C, Metz CE, Doi K. Sabeti V. Solomon SL. Digital radiography of subtle pulmonary abnormalities: an ROC study of the effect of pixel size on observer performance. Radiology 1986; 158:21-26. Bramble JM, Huang HK, Murphy MD. Image data compression. Invest Radiol 1988; 23:707-712. Elnahas S. Jost RG, Dunham JG. Compression of digital diagnostic images in ra-

January

1991

1.G

z

z 0

0

I0

I0

I-

U.

w

I

0.6

> ICo

0

cc

UNCOMPRESSED

U.

w

/NNAGE

w

> IC/)

0

jc,

,

cc U.

z 0

0.8

> I-. U)

0.4

0

0

50.1

0.

0.

w

w

cc

cc

0.

w

0.2

cc I-

I-

0.2

I

FALSE

POSITIVE

0.4

FALSE

FRACTION

0.6

POSITIVE

0.8

1.0 FALSE

FRACTION

8. Figures 25:1

and

7-9.

ROC

50:1

compressed

I-.

curves

compressed

indicating images.

the relative (9) Composite

accuracy ROC

curves

of interstitial indicating

infiltrate (7) and overall diagnostic

and 50:1

images.

1

diotogy.

Proceedings

ference

on Computer

diology.

w

4.

:D -I

> 5.

-J

w 6.

0 I-

z

w

7.

0 Ll LI

w 0 C-)

8.

w 9.

0

LJ

101

SPATIAL Average Fourier in two-dimensional

spectra

FREQUENCY for compressed frequency space

dial spatial frequency axis) to obtain an estimate image. Note the gradual decrease at high spatial

(CYCLES/MM) and were

uncompressed averaged

over

10. chest images. Fouriall angles (at the ma-

of the average frequency content of each frequencies as the compression ratio in-

1 1.

creases. 12.

Press,

178

Number

#{149}

1

Reston,

of the Va:

Eighth

Applications American

Con-

in RaCollege

of

Radiology, 1984; 383-409. Lo S-C, Huang HK. Compression of radiological images with 512, 1,024, and 2,048 matmices. Radiology 1986; 161:519525. Chen W, Pratt W. Scene adaptive codes. IEEE Trans Biomed Eng 1984; COM32:225-232. Yonekawa H, Ishimitsu Y, Akune J, Takeuchi H. Adaptive block cosine transform coding using a new quantization algorithm. Proc SPIE 1989; 1093:474-485. Ishigaki T, Sakuma 5, Ikeda M, Itoh Y, Suzuki M, Iwai S. Clinical evaluation of irreversible image compression: analysis of chest imaging with computed radiography. Radiology 1990; 175:739-743. Ishida M, Kato H, Doi K, Frank PH. Devebopment of a new digital radiographic image processing system. Proc SPIE 1982; 347:42-48. MacMahon H, Metz CE, Doi K, Kim T, Giger ML, Chan HP. Digital chest radiogmaphy: effect on diagnostic accuracy of hard copy, conventional video, and reversed gray scale video display formats. Radiology 1988; 168:669-673. Metz CE. Some practical issues of experimental design and data analysis in radiological ROC studies. Invest Radiol 1989; 24:234-245. Metz CE. ROC methodology in radiologic imaging. Invest Radiol 1986; 21:720733. Swets JA, Pickett RM. Evaluation of diagnostic systems: methods from signal detection

Volume

FRACTION

bone lesion (8) detection with uncompressed accuracy with uncompressed and 25:1 and

z

Figure 10. em coefficients

POSITIVE

9.

theory.

New

York:

Academic

1982.

Radiology

#{149} 179