Downloaded from www.ajronline.org by 121.204.21.102 on 10/14/15 from IP address 121.204.21.102. Copyright ARRS. For personal use only; all rights reserved
467
Detection Disease: Digital
Kathleen M. Buckley1 Cornelia M. Schaefer1’2
Reginald
Greene1
Steve
Agatston1 Jane Fay1 Henry J. Llewellyn1 Helen E. Mrose1 Jami R. Rubens1
of Bullous Lung Conventional Radiography vs
Storage
The detection of radiologic diagnosis contrast resolution five postprocessing
Phosphor
Radiography
fine linear contours and altered aeration are requirements for the of bullous lung disease and are severe measures of the spatial and of chest imaging systems. We compared plain film radiography with algorithms of storage phosphor digital radiography (2144 x 1744 x
10 bit matrix with 0.2-mm pixel size) in the detection of CT-proved bullous lung disease (35 patients and 25 normal control subjects). Receiver-operating-characteristic analysis of 2160 observations by six interpreters was done to evaluate the observers’ performance. By analysis of variance (p < .05), we found that the default digital algorithm and the three edge-enhancing algorithms of high and medium frequencies performed less well than plain films did, but the differences fell short of statistical significance. Gray-
scale reversal
was the only digital algorithm
that
performed
significantly
less
well
than
plain films did.
We conclude that any differences between digital algorithms and plain films in the detection of bullous lung disease were too small to be detected in this study. Any difference between the two methods in providing clinically important, diagnostic information AJR
is likely to be insignificant.
March 1991
156:467-470,
Storage phosphor digital radiography has a number of technical advantages (e.g., high dynamic range, flexible adjustment of contrast and density scale, and other postprocessing capabilities) that may offset its lower spatial resolution relative to plain film radiography [i -6]. The detection of fine linear contours and altered aeration that is necessary for the diagnosis of bullous lung disease provides a unique test of the spatial and contrast resolution of an imaging system. Previous experimental and clinical studies have reported discordant results in the detection of fine linear structures with digital radiography and have suggested the important role of edge-enhancement postprocessing [2-4, 7, 8]. We compared plain film radiography with five postprocessing algorithms of storage phosphor radiography in the detection of bullous lung disease. Materials Received sion October
March 29, 1990; accepted 10, 1990.
after
revi-
Department of Radiology, Massachusetts General Hospital and Harvard Medical School, 32 Fruit St., Boston, MA 02114. Address reprint requests to R. Greene. 2Presont address: Department of Radiology, 1
Medizinische Germany.
Hochschule
Hannover,
0361-803x/91/1563-0467 © American Roentgen Ray Society
Hannover,
and Methods
We studied a total of 60 patients who had undergone CT of the chest for clinical indications. Thirty-five patients had a diagnosis of bullous lung disease. Twenty-five persons served control subjects. There were 40 men and 20 women from 25 to 88 years of age (mean standard error = 60 ± 1 .5 years). Each patient in the study group had posteroanterior radiographs
of the
chest
obtained
with
storage
phosphor
and
conventional
technique
as ±
within
1 week of the CT examination ofthe thorax. The CT examinations (GE 9800, General Electric Medical Systems, Milwaukee, WI) consisted of multislice scans (iO mm thick at i 0-mm intervals) through the chest with selected thin sections (1.5 mm thick with high-resolution algorithm) as warranted by clinical circumstances. The official interpretation of the CT findings served as the basis for the expected findings on the radiographs. Criteria for the diagnosis of bullous
lung
disease
included
the identification
of discrete,
localized
areas
of low attenuation
BUCKLEY
468
surrounded pleura
by well-defined,
or compressed
bullae
were
they
were
surrounded located
of localized
Downloaded from www.ajronline.org by 121.204.21.102 on 10/14/15 from IP address 121.204.21.102. Copyright ARRS. For personal use only; all rights reserved
were
classes
diameter
(n
diffuse
to size:
and
eight
2-4
cm
subjects
in most
were
than
had
and
with
with
factor,
same
and
exposure
nominal
and
1 2:1
Erlangen,
radiographs
1 40
kVp,
oscillating
phototimed
Systems,
digital
factors:
focus,
tance,
average
skin
entrance
with
both
techniques
with
Lanex
medium
with
standard
The OC
(Kodak
digitized
America,
with
Tustin,
a helium-neon CA).
A laser
mR
and
processor,
laser printer
x
106
were
automatic
film
bit
(ST,
Fuji,
readout
unit
NY) size)
(TCR
201
area
phosphor
digital were
radiographs
with
images
were
processed
enhanced
with
an unsharp
to the
A + f (A
image,
unsharp
=
following
formula:
B)
-
produced
by
the final
The
moderate,
The
4) and five
Mode
at
ance
were
degree
Toshiba
enhancement
17 x 21 cm hardcopies
enhanced
Hurter
of
the
f
mask
enhancement
=
(B) was
produced
(cycles/mm)
edge-enhanced edge
image
enhancement
effect
(A
was
(mode
and high
of gray-scale
pronounced
that
was
image
was
B) to the
-
determined
by
were
spatial range
(mode
of high
2) and
(modes
1,
as follows:
algorithm
Driffield
reversal
enhancement
5) frequencies.
conventional
of
This
of edge
the
default
frequency
image,
frequency
(A*).
algorithms
The
original unsharp
factor (f), which could be chosen on a scale of
medium
of an optimal
contrast
spatial
addition
and
digital
1:
=
The
image
weighted
image.
A
mask.
pixel values across a certain area. The size of
determined in the
3, and
Japan) ,
B
enhanced
subtle,
processing
Kanagawa,
according
the flexible weighting 0 to 1 6. We tested
obtained
matrix
Edges
technique
edge-enhanced and
original
C/kg)
Rochester,
x 10
produced
dis-
measurements, (5.16
images
film,
plates
phosphor
1 .2-mm
source-to-image
M7B
acquired
the
2 MP, Siemens Medical
conventional
x 2144
storage
20
with
filter,
to previous
was
36#{176}C. The digital images (1744 with
71
According
screens,
chemicals
1 83-cm
storage
algorithms.
by locally averaging
of diffuse
obtained
aluminum
(Tridoros
exposure
[9].
were
3-mm
grid,
exposure Germany).
the
and
A*
this
conventional
to five
subtraction
A*
emphysema. The
of the
at 36.5#{176}C. The
1991
axis
patients
a CT diagnosis
according mask
Japan)
chemicals
March
or
22);
=
Kanagawa,
into
2 cm in short (n
AJA:156,
standard
causes
divided
7). Seven
=
the cases
Other
in diameter
(n
cases
bronchiectasis,
bullae less
in diameter
control
lung,
The
) small,
6); (2) intermediate,
=
bullae
(1
but
(Fugi,
of inverted
In some
parenchyma.
honeycomb
excluded.
(3) large, greater than 4 cm large
normal
(e.g.,
specifically
consisting
or both. emphysema,
otherwise
according
margins
parenchyma
by
within
hyperinflation
oligemia) three
paper-thin
lung
ET AL.
(SR-i) radiograph.
(H&D)
curve
frequencies centered
approximated It used with
slight
(enhancement
the
appear-
a sigmoid,
long-
nonlinear
edge
factor,
0.5;
at 1 .4 cycles/mm).
Fig. i.-A-G, cT scan (A), plain film (B), and five different algorithms (C-G) of a storage phosphor digital radiograph show a left-sided intraparenchymal cyst (arrows). Five algorithms are default mode (C), gray-scale reversed mode (D), moderately edge-enhanced mode of high frequencies (E), markedly
edge-enhanced
mode of high fre-
quencies (F), and moderately edge-enhanced mode of medium frequencies (G). Notice that edge enhancement of high and medium frequencies (E, F, and G) amplifies surrounding fine linear rim as compared with plain film (B) and default digital mode (C).
A
D
B
E
C
F
G
AJA:i56,
March
BULLOUS
1991
Mode 2: The gray-scale reversed H&D curve as the default algorithm
algorithm
(SR-2)
used
LUNG
an identical
except for reversal of gray-scale
Downloaded from www.ajronline.org by 121.204.21.102 on 10/14/15 from IP address 121.204.21.102. Copyright ARRS. For personal use only; all rights reserved
polarity.
Mode 3: The edge-enhanced algorithm (SR-3) performed a moderate enhancement of high spatial frequencies and was identical to the default algorithm except for a larger enhancement factor of 3. Mode 4: The edge-enhanced algorithm (SR-4) performed a marked enhancement
of
high
spatial
frequencies
and
was
identical
to
the
default mode 1 except for a larger enhancement factor of 5. Mode 5: The edge-enhanced algorithm (SR-5) performed a moderate enhancement of medium spatial frequencies and was identical to mode 3 except for the enhanced frequency range (enhancement factor 3, enhanced frequency range centered at 0.71 cycles/mm, Figs. iA-iG). The conventional and digital images were intermingled in random order and divided into three subsets. Six board-certified radiologists interpreted the subsets in differing order to avoid interpreter bias. Three separate sessions per interpreter extended over a 4-week period.
The
definitions
and
exclusions
in this study were reviewed not
know
the
of bullous
lung
disease
of
normal
to
abnormal
cases.
They
did were
asked to indicate the presence of bullae according to a five-level scale of confidence (1 definitely negative, 2 probably negative, 3 indeterminate, 4 probably positive, 5 definitely positive). =
=
=
=
Interpretation
time
was
(ROC)
interpreter obtained
data by
for
digital algorithm that performed significantly worse than plain films did (Fig. 2 and Table 1). The algorithm with moderate enhancement of high spatial frequencies (mode 3) performed best among the digital modes. Five of the six interpreters had higher ROC areas with the plain films than with the best digital algorithm. However, the differences were not statistically significant (Table 2). Analysis of performance according to the size of bullae indicated that the gray-scale reversed mode was inferior to plain films in the detection of small and intermediate-sized bullae. Plain films and the best digital algorithm with moderate enhancement of high spatial frequencies (mode 3) were equivalent in the detection of bullae of all sizes (Table 3).
10
z 08
0 0
=
not
Observer performance characteristic
limited.
was tested according
analysis a total
of
of 21
pooled,
60 observations.
a maximum-likelihood
w >
to receiver-operating-
averaged,
and
individual
algorithm,
and
06
U)
ROC curves were
curve-fitting
U-
per-
0 0
w 04
ceptual accuracy was measured by the area under the ROC curve 110-13]. ROC curves were produced by fitting the pooled interpreter data from
all observers
and
ically [1 0]. Averaged by
individually
fitting
were
used
to demonstrate
ROC areas (A7 the
performance
± standard
confidence
error)
ratings
for
were
each
graph02
produced
observer
and
averaging the estimated areas across the observers. These areas were used to measure performance of each imaging mode [i 1-13]. The significance of differences between the A of the six imaging modes was tested with a one-factor analysis of variance for repeated measures. Paired A values were compared with Scheff#{233}’s F-test at a two-tailed p less than .05 [i4]. A preliminary arc-sine transformation was
performed
homogeneity
to stabilize
the
normal
distribution
and
data
and
to meet
of variances
the
assumptions
needed
of
for analysis
of variance Ii 4]. The arc-sine transformation was considered the appropriate procedure because it applies to variables that express proportions. Performance with respect to the three subclasses of bullae (small, intermediate, and large) was tested with pooled and individual data for
plain
film
performance.
and
for
the
Significance
digital of
algorithms differences
with for
the the
best pooled
and
worst
data
that
incorporated
estimations
of variances
and
00 0.0 Fig.
02
06
04
2.-Receiver-operating-characteristic
covariances
of the ROC parameters into a univariate z-score test (p < .05) for the true-positive fraction at a constant false-positive fraction of 0.2 [1 6-i 8].
high spatial
TABLE
curves
Performance with all digital algorithms (averaged ROC area ± standard error) was lower than that with plain films. However, no statistically significant difference was found between plain films and four of the five digital algorithms at a p level of .05. The gray-scale reversed algorithm was the only
10
derived
from
pooled
frequencies.
1: Average
Performance
Operating-Characteristic
and Range
(ROC)
Areas
Plain
of Receiver-
by Six Readers
ROC Area
Technique/
Postprocessing
Algorithm
film
Mean 0.76
SR-i, Default: slight edge enhancement (high
± SEM
Range
Significance: VS Plain
SR
Film
± 0.04 ± 0.02
0.64-0.87 0.67-0.75
p < .19
0.72
± 0.05 ± 0.02
0.37-0.72 0.67-0.80
p < .05 p < .31
0.71
±
0.02
0.60-0.80
p