usefulness of computerized method for lung nodule detection in digital chest radiographs using...
TRANSCRIPT
Usefulness of Computerized Methodfor Lung Nodule Detection in DigitalChest Radiographs Using Temporal
Subtraction Images
Takatoshi Aoki, MD, Nobuhiro Oda, PhD, Yoshiko Yamashita, MD,Keiji Yamamoto, PhD, Yukunori Korogi, MD
Ac
FrEnKiRa(NMNto
ªdo
10
Rationale andObjectives: The aim of this studywas to evaluate the usefulness of a novel computerizedmethod for lung nodule detectionon digital chest radiographs using temporal subtraction images.
Materials and Methods: To significantly reduce the number of false-positive results while maintaining high sensitivity, temporal subtrac-
tion images, which can enhance interval changes on sequential chest radiographs, were used. Fifty-one caseswith lung nodules <3 cmand51 cases without lung nodules were selected for an observer performance test. Twelve radiologists participated in this observer perfor-
mance test. The radiologists’ performance was evaluated using receiver-operating characteristic analysis, on a continuous rating scale.
To estimate the numbers of cases affected beneficially and those affected detrimentally using this computerized method, the computer
output was assumed to have an effect on an observer’s diagnosis when there was a difference in rating score of $30% between the firstand second ratings.
Results: The average area under the curve for all radiologists increased significantly from 0.849 to 0.950 with the computerized method
(P < .001). The mean number of cases affected beneficially was significantly higher than that of cases affected detrimentally (8.92 vs 1.25,P < .001).
Conclusions: The novel computerized method using temporal subtraction images would be useful in detecting lung nodules on digital
chest radiographs.
Key Words: Lung nodule; lung cancer; computer-aided-diagnosis; chest radiograph.
ªAUR, 2011
Although chest radiography is the most prevalent
screening procedure for detecting lung lesions, because
it is economical and easy to use, radiologists are some-
times unaware of lung nodules that are perceptible on chest
radiographs in retrospect (1,2), and it is even more difficult to
detect small lung nodules (3). Computer-aided diagnosis
(CAD) is now considered an approach that might improve
the efficacy of radiologic image interpretation. For the detec-
tion of subtle lung nodules, however, CAD of a single chest
radiograph suffers from interference by anatomic noise, leading
to low specificity for an acceptable level of sensitivity (4). The
ad Radiol 2011; 18:1000–1005
om the Department of Radiology, University of Occupational andvironmental Health, School of Medicine, Iseigaoka 1-1, Yahatanishi-ku,takyushu 807-8555, Japan (T.A., Y.Y., Y.K.); the Department ofdiological Technology, Kyoto College of Medical Science, Kyoto, Japan.O.); and the Information Network System Department, Kansai Division,itsubishi Space Software Company, Ltd, Hyogo, Japan (K.Y.). Receivedovember 11, 2010; accepted March 22, 2011. Address correspondence: T.A. e-mail: [email protected]
AUR, 2011i:10.1016/j.acra.2011.04.008
00
temporal subtraction technique is a CAD method in which
a previous chest radiograph is subtracted from a current radio-
graph so that interval changes are enhanced. Several studies
have found that this technique improves the diagnostic accu-
racy of newly formed lung nodule (5–7). However,
misregistration artifacts of this system, due to mismatching of
normal anatomic structures in current and previous images,
cause false-positive findings and sometimes lead tomisinterpre-
tation. In addition, both previous and current images are
required for this CAD system, which may limit its clinical use.
We are developing a novel computerized method to assist
radiologists in the detection of lung nodules by using temporal
subtraction images. Our purpose in this study was to evaluate
the usefulness of this computerized method on radiologists’
performance for lung nodule detection.
MATERIALS AND METHODS
Scanning Protocol
All chest radiographs were exposed at 100 kV with a 10:1 grid
and were obtained using a computed radiographic system
Academic Radiology, Vol 18, No 8, August 2011 CAD FOR LUNG NODULE IN DIGITAL RADIOGRAPHS
(FCR; Fuji Photo Film, Tokyo, Japan). The imaging plate
(model ST-V; Fuji Photo Film) was 35 � 35 cm (matrix
size, 1760 � 1760; gray level, 10 bits; pixel size, 0.2 mm).
Hospital institutional review board approval was obtained,
and informed consent was not required by the institutional
review board.
Figure 1. Overall scheme of computerized method for lung noduledetection.
Computerized Scheme
A diagram of our computerized scheme for lung nodule
detection is shown in Figure 1 (8). To significantly reduce
the number of false-positive findings while maintaining high
sensitivity, we used temporal subtraction images produced
using a commercially available temporal subtraction system
(Truedia/XR; Mitsubishi Space Software, Amagasaki, Japan)
(9,10). In this system, a registration algorithm previously
described (9) was used. First, extraction of the chest region
was applied to ribcage edge detection on chest radiographs
(11). The CAD system was applied to correct the background
density trend on chest subtraction image (12). Then, nodule
candidates were extracted using multilevel thresholding with
various pixel values, which were determined from the area
of the histogram of pixel values on the temporal subtraction
image (13). Twenty-one image features for each nodule candi-
dates were obtained from both subtraction images and current
chest radiographs. The image features on nodules included
three morphologic features (area, circularity, and irregularity)
obtained from the subtraction image and nine gray-level
features (mean, maximum, minimum, variance, contrast,
skewness, kurtosis, energy, and entropy) obtained from histo-
graphic analysis of pixel values within the nodule on both
subtraction and current images. A rule-based approach and
aMahalanobis distance discriminant analysis with 21morpho-
logic and gray-level features were applied to remove false-
positive candidates corresponding to non-nodules (14,15).
Finally, the computer output was superimposed on the
detected nodules on current chest radiographs using markers
(plus signs; Fig 2).
Performance of the CAD System
This CAD system was first preliminarily applied to a database
consisting of 88 radiographswith 90 lung nodules. The nodule
presence and absence information is provided by the database.
Two board-certified radiologists independently classified the
radiographs with lung nodules according to the subtlety of
each lung nodule using a five-point scale: 1 = extremely
subtle, 2 = very subtle, 3 = subtle, 4 = relatively obvious,
and 5 = obvious. The evaluationswere based on both the orig-
inal image and the CAD output. The subtlety of the nodules
was calibrated beforehand using a large number of training
cases included in a digital image database for chest radiographs,
which is publicly available (16). Among the 90 lung nodules,
there were 21 (23%) with subtlety scores of 1, 31 (35%) with
subtlety scores of 2, 23 (26%) with subtlety scores of 3, 13
(14%) with subtlety scores of 4, and two (2%) with subtlety
scores of 5. Using this database with a high degree of subtlety,
we achieved detection sensitivity of 80%, with an average of
2.6 false-positive detections per image.
Observer Performance Study
We selected 51 cases with lung nodules <3 cm (17 lung carci-
nomas, 17 metastatic lung tumors, and 17 benign nodules) and
51 cases without lung nodules on the basis of computed tomo-
graphic confirmation for an observer performance test of lung
nodule detection. Calcified lung nodules were excluded in
this study. The subtlety ratings of the nodules were classified
by two board-certified radiologists in the same way as for
the classification of the database. There were 15 (29%) with
subtlety scores of 1, 26 (51%) with subtlety scores of 2, five
(10%) with subtlety scores of 3, five (10%) with subtlety scores
of 4, and none with subtlety scores of 5.
Assessment of each radiologist’s diagnostic accuracy was
determined using receiver-operating characteristic analysis
with a continuous rating scale. Twelve radiologists, including
four attending radiologists with$10 years of experience, four
radiology fellows with 4 to 6 years of experience, and four
radiology residents with <2 years of experience, participated
in this observer performance test. The sequential test method
was used in the observer performance study (17). Confidence
level regarding the presence or absence of lung nodules was
marked above both ends of the 7-cm line using a black ball-
point pen on the line for the first rating (without CAD).
Only current radiographs were displayed on the 1536 �2048 line, 20.8-inch (diagonal), liquid crystal display monitor
(Dome C3; Planar Systems, Inc, Beaverton, OR). Neither
previous radiographs nor subtraction images were displayed.
Then, the current image with the computer output was dis-
played on the liquid crystal display monitor. The radiologists
could freely remove the markers indicating nodule candidates
or change the sizes of the markers during the observer perfor-
mance test. Observers were allowed to change their first
ratings using a red ballpoint pen on a same line, if the second
ratings were different from the first ones (with CAD). Before
the test, observers were shown 20 training cases (10 cases with
1001
Figure 2. Lung carcinoma in a 68-year-old
man. (a) Current chest image using for input
image. (b) Previous chest image. (c) Temporal
subtraction image using for input image.(d) Computer-aided diagnosis system output
image. Two suspected areas indicated by small
plus signs. One area contains an actual nodule(arrow), and the other contains a false-positive
finding (arrowhead).
AOKI ET AL Academic Radiology, Vol 18, No 8, August 2011
normal findings and 10 nodule cases), so that they could learn
the rating method and how to take into account the computer
output in their decisions. The performance of the CAD
system and clinical parameters (patients’ ages and sex) were
not indicated. Radiographs were presented randomly, and
the reading time was not limited.
To estimate the numbers of cases affected beneficially and
those affected detrimentally by this computerized method,
we assumed that the candidate nodules of the CAD system
had an effect on an observer’s diagnosis when there was
a difference in rating score of $30% between the first and
second ratings.
Statistical Analysis
Areas under the receiver-operating characteristic curve (Az)
were computed using the computer program LABROC5,
provided by Metz et al (18). The statistical significance of
the difference in Az values without and with CAD output
images was estimated using the Dorfman-Berbaum-Metz
method (19). Comparisons of the improvement in Az values
among three observer categories were performed using one-
way analysis of variance. If a significant difference was found,
Bonferroni’s correction was used for post hoc analysis. Using
Bonferroni’s correction, we set the significance level at
1002
P < .0033 to adjust for multiple comparisons. The difference
between the average number of cases affected beneficially and
those affected detrimentally by CAD output images was
analyzed using a two-tailed paired Student’s t test. All statistical
analyses were performed using StatView version 5.0 (SAS
Institute Inc, Cary, NC).
RESULTS
The Az values without and with the CAD for each observer
are summarized in Table 1. All of the radiologists achieved
improved diagnostic performance when the computer output
was available. The average Az value for all radiologists
increased significantly from 0.849 to 0.950 with the comput-
erized method (P < .001; Fig 3). All three observer categories
showed significant improvements in diagnostic accuracy with
the CAD system (P < .001). The difference in the level of
improvement in diagnostic accuracy with the CAD system
was significant among the three observer categories (P <
.001). After Bonferroni’s correction was applied, the improve-
ments in the average Az values were significantly greater for
the radiology residents than for the attending radiologists
(P < .001).
The numbers of cases with clinically relevant changes
in confidence levels for each observer are shown in
TABLE 1. Comparison of AUC Values without and with CAD
Observer
AUC
PWithout CAD With CAD
Residents
A 0.814 0.961
B 0.826 0.959
C 0.768 0.927
D 0.759 0.874
Average 0.792 0.930 <.001
Fellows
E 0.853 0.951
F 0.895 0.952
G 0.799 0.937
H 0.812 0.934
Average 0.840 0.944 <.001
Attending radiologists
I 0.931 0.982
J 0.903 0.970
K 0.933 0.972
L 0.900 0.979
Average 0.917 0.976 <.001
Average total 0.849 0.950 <.001
AUC, area under the curve; CAD, computer-aided diagnosis.
Figure 3. Comparison of average receiver-operating characteristic
curves of all radiologists for lung nodule detection without and withthe computer-aided diagnosis (CAD) system. The average area
under the curve (Az) for all radiologists increased significantly from
0.849 to 0.950 with the CAD system (P < .001, Dorfman-Berbaum-
Metz method [19]).
Figure 4. Bar graph showing the number of cases affected by the
computer output in confidence levels with regard to lung nodulecases. The average number of cases affected beneficially (4.0) was
significantly larger than the number affected detrimentally (0.58)
(P < .001, two-tailed paired t test).
Figure 5. Bar graph showing the number of cases affected by thecomputer output in confidence levels with regard to healthy cases.
The average number of cases affected beneficially (4.9) was signifi-
cantly larger than the number affected detrimentally (0.7) (P = .002,
two-tailed paired t test).
Academic Radiology, Vol 18, No 8, August 2011 CAD FOR LUNG NODULE IN DIGITAL RADIOGRAPHS
Figures 4 and 5. Among the 51 lung nodule cases, the average
number of cases affected beneficially (4.0) was significantly
larger than the number of cases affected detrimentally (0.6)
(P < .001; Fig 4). Among the 51 healthy cases, the average
number of cases affected beneficially (4.9) was significantly
larger than the number of cases affected detrimentally (0.7)
(P = .002; Fig 5).
DISCUSSION
Even though computed tomographic techniques continue to
develop, chest radiography remains the first and the most
common examination for the detection of lung nodules.
However, it has been shown that detecting small lung lesions
on chest radiographs is a difficult task for radiologists (1,2).
Currently, CAD technology has a complementary role in
clinical practice as a second opinion. Given the increasingly
digital nature of chest radiography, CAD will most certainly
be an integral part of clinical practice. Freedman et al (20)
reported positive findings from a clinical trial for an indepen-
dent validation test with 80 lung cancer cases with nodules 7
to 30 mm in diameter and 160 cases without nodules using of
a commercial CAD system, achieving 66% nodule detection
sensitivity with an average of 5.3 false-positive findings per
image. However, the current standard of CAD’s performing
a supporting role for lung nodule detection is rather limited.
CAD of a single chest radiograph is impeded by normal
anatomic structures, resulting in low specificity for a satisfac-
tory level of sensitivity. Consequently, CAD is expected to
1003
AOKI ET AL Academic Radiology, Vol 18, No 8, August 2011
continue to evolve to meet the increasing new challenges of
chest radiography.
The emergence of digital chest radiography in the 1980s led
to new techniques to improve the detection of subtle pulmo-
nary lesions. The temporal subtraction technique, which
selectively highlights areas of interval change by subtracting
the patient’s previous radiograph from the current one, is
one notable approach. Generally, areas that stand out in the
uniform gray background indicate interval changes. The
fact that temporal subtraction helps radiologists discern subtle
lung abnormalities, including nodules, is supported by several
studies (5–7). At the same time, the image quality of the
difference image strongly depends on the two-dimensional
registration of the two radiographs. Misregistration artifacts
are caused by mismatching normal anatomic structures in
current and previous images and are due mostly to the ribs,
diaphragm, clavicle, and blood vessels. The temporal subtrac-
tion system cannot indicate nodule candidate directly, so that
radiologists must decide whether the standout area on the
image is a true-positive finding or artifact. These misregistra-
tion artifacts cause false-positive findings and sometimes lead
to misinterpretation. Thus, an automated lung nodule detec-
tion method using temporal subtraction is expected to help
radiologists.
Because of the use of this CAD system, significant improve-
ments in diagnostic accuracy for pulmonary nodule detection
were experienced by all 12 radiologists in the observer perfor-
mance study. Although all observer categories benefited from
this CAD system, those with less experience tended to benefit
more. This evidencemay indicate that the radiologists’ perfor-
mance in the detection of lung nodule depended on their
experience. Radiologists need some experience in nodule
detection based on the temporal subtraction system, but the
use of CAD may reduce the impact of traditional experience.
For nodule and healthy cases, we analyzed the influence of
CAD output images separately. The CAD output affected
observer confidence in terms of both beneficial and detri-
mental effects in both cases. The average number of cases
affected beneficially by computer output was significantly
larger than the number of cases affected detrimentally, even
in healthy cases. The use of this CAD system can decrease
false-positive findings and may lead to avoidance of unneces-
sary further examinations, such as chest computed
tomography.
Temporal subtraction may be most practical in follow-up
examinations of patients with a high risk for lung tumors,
such as patients with interstitial pneumonia, pulmonary
emphysema, and extrapulmonary malignant neoplasm. Chest
radiography is most commonly used for follow-up and is very
important for lung tumor detection in these patients.
Although considerable attention has recently been given to
the detection of early-stage lung cancer through low-dose
computed tomographic screening of individuals at high risk
(21,22), chest radiography is still the preferred imaging
modality used for lung nodule detection. Also, this CAD
system may be useful in repeated lung cancer screening
1004
using digital chest radiographs and in assisting physicians
who do not have easy access to computed tomographic
scanners in their daily practice.
Our study had several limitations. First, we used our own
image database for chest radiographs with lung nodules.
Although we achieved detection sensitivity of 80%, with an
average of 2.6 false-positive detections per image using the
database containing many subtle nodules, our results cannot
be compared directly to those obtained in other previous
studies, because of the differences in the database. Further
studies using common CAD testing databases may be needed
to validate the performance of our CAD system. Second, time
spent reviewing interpretations without and with this CAD
system was not recorded for comparison, even though this
study showed that the new CAD system, using temporal
subtraction images, significantly improved diagnostic accu-
racy for lung nodule detection. Radiologists’ performance
in daily practice is affected not only by diagnostic accuracy
but also by interpretation time. Comparison of reviewing
time in the interpretation of radiographs without and with
this CAD systemmay also be important to evaluate the useful-
ness of this system on radiologists’ performance for lung
nodule detection. Third, this study was retrospective. Further
prospective study of a greater number of cases is likely neces-
sary to confirm the clinical usefulness of this computerized
method.
In conclusion, although the new computerized method
using temporal subtraction images has its limitations with
regard to clinical use, it can assist in subtle lung nodule detec-
tion on digital chest radiographs.
REFERENCES
1. Austin JH, Romney BM, Goldmith LS. Missed bronchogenic carcinoma:
radiographic findings in 27 patients with a potentially resectable lesion
evident in retrospect. Radiology 1992; 182:115–122.
2. Muhm JR, Miller WE, Fontana RS, et al. Lung cancer detected during
a screening program using 4-month chest radiographs. Radiology 1983;
148:609–615.
3. Heelan RT, Flehinger BJ, Melamed MR, et al. Non-small-cell lung cancer:
results of the New York Screening Program. Radiology 1984; 151:
289–293.
4. Samei E, Flynn MJ, Eyler WR. Detection of subtle lung nodules: relative
influence of quantum and anatomical noise on chest radiographs. Radi-
ology 1999; 213:727–734.
5. Uozumi T, Nakamura K, Watanabe H, et al. ROC analysis of detection of
metastatic pulmonary nodules on digital chest radiographs with temporal
subtraction. Acad Radiol 2001; 8:871–878.
6. Johkoh T, Kozuka T, Tomiyama N, et al. Temporal subtraction for detec-
tion of solitary pulmonary nodules on chest radiographs: evaluation of
a commercially available computer-aided diagnosis system. Radiology
2002; 223:806–811.
7. Kakeda S, Nakamura K, Kamada K, et al. Improved detection of lung
nodules by using a temporal subtraction technique. Radiology 2002;
224:145–151.
8. Oda N, Kido S, Shono H, et al. Development of computerized
system for detection of pulmonary nodules on digital chest radio-
graphs using subtraction images. Jpn J IEICE Trans 2004; 87D:
208–218.
9. Kano A, Doi K, MacMahon H, et al. Digital image subtraction of temporally
sequential chest images for detection of interval change. Med Phys 1994;
21:453–461.
Academic Radiology, Vol 18, No 8, August 2011 CAD FOR LUNG NODULE IN DIGITAL RADIOGRAPHS
10. Ishida T, Katsuragawa S, Nakamura K, et al. Iterative image warping tech-
nique for temporal subtraction of sequential chest radiographs to detect
interval change. Med Phys 1999; 26:1320–1329.
11. Xu XW, Doi K. Image feature analysis for computer-aided diagnosis: accu-
rate determination of ribcage boundary in chest radiographs. Med Phys
1995; 22:617–626.
12. Katsuragawa S, Doi K, MacMahon H. Image feature analysis and
computer-aided diagnosis in digital radiography: detection and character-
ization of interstitial lung disease in digital chest radiographs. Med Phys
1988; 15:311–319.
13. Xu XW, Doi K, Kobayashi T, et al. Development of an improved CAD
scheme for automated detection of lung nodules in digital chest images.
Med Phys 1997; 24:1395–1403.
14. Duda RO, Hart RE, Stork DG. Pattern Classification. 2nd ed. New York:
John Wiley, 2001.
15. Jain AK, Duin PW, Mao J. Statistical pattern recognition; a review. IEEE
Trans Pattern Anal Mach Intell 2000; 22:17–24.
16. Shiraishi J, Katsuragawa S, Ikezoe J, et al. Development of a digital image
database for chest radiographs with and without a lung nodule: receiver
operating characteristic analysis of radiologists’ detection of pulmonary
nodules. AJR Am J Roentgenol 2000; 174:71–74.
17. Kobayashi T, Xu XW, MacMahon H, et al. Effect of a computer-aided diag-
nosis scheme on radiologists’ performance in detection of lung nodules on
radiographs. Radiology 1996; 199:843–848.
18. Metz CE, HermanBA, Shen JH.Maximum-likelihood estimation of receiver
operating characteristic (ROC) curves from continuously-distributed data.
Stat Med 1998; 17:1033–1053.
19. Dorfman DD, Berbaum KS, Metz CE. ROC rating analysis: generalization
to the population of readers and cases with the jackknife method. Invest
Radiol 1992; 27:723–731.
20. Freedman MT, Osicka T, Lo SCB, et al. Methods for identifying changes in
radiologists’ behavioral operating point of sensitivity-specificity tradeoffs
within an ROC study of the use of computeraided detection of lung cancer.
Prop SPIE 2001; 4324:184–194.
21. Henschke CI, Naidich DP, Yankelevitz DF, et al. Early lung cancer action
project: initial findings on repeat screenings. Cancer 2001; 92:153–159.
22. Patz EF, Black WC, Goodman PC. CT screening for lung cancer: not ready
for routine practice. Radiology 2001; 221:587–591.
1005