DOI: 10.1148/radiol.2231010344
(Radiology 2002;223:199-203.)
© RSNA, 2002
Detection of Lung Nodules on Digital Chest Radiographs: Potential Usefulness of a New Contralateral Subtraction Technique1
Shunji Tsukuda, MD,
Atsuko Heshiki, MD,
Shigehiko Katsuragawa, PhD,
Qiang Li, PhD,
Heber MacMahon, MD and
Kunio Doi, PhD
1 From the Department of Radiology, Saitama Medical School, Saitama, Japan (S.T., A.H.); and Kurt Rossmann Laboratories for Radiologic Image Research, Department of Radiology, University of Chicago, 5481 S Maryland Ave, Chicago, IL 60637. From the 1999 RSNA scientific assembly. Received February 6, 2001; revision requested March 1; revision received August 20; accepted September 20. Supported by U.S. Public Health Service grants CA62625 and CA64370. Address correspondence to K.D. (e-mail: k-doi@uchicago.edu).
 |
ABSTRACT
|
|---|
PURPOSE: To evaluate the potential usefulness of a contralateral subtraction technique developed for radiologists’ performance in the detection of subtle lung nodules on chest radiographs.
MATERIALS AND METHODS: Fifty chest radiographs (25 normal and 25 abnormal with a subtle lung nodule) that were digitized with a 0.175-mm pixel size and 4,096 gray levels were used. Twelve radiologists (10 attending and two residents) participated in observer tests and read both original and contralateral subtraction images with a sequential testing method. Radiologists’ performance was evaluated by means of receiver operating characteristic analysis with use of a continuous rating scale. The beneficial and detrimental effects of the contralateral subtraction technique on the radiologists’ performance were also evaluated.
RESULTS: The area under the receiver operating characteristic curve values obtained without and with contralateral subtraction images were 0.926 and 0.962, respectively. Results indicated that the contralateral subtraction images significantly (P < .05) improved diagnostic accuracy, particularly for radiologists with limited experience.
CONCLUSION: The contralateral subtraction technique can assist radiologists in the correct identification of subtle lung nodules on chest radiographs.
© RSNA, 2002
Index terms: Computers, diagnostic aid • Diagnostic radiology, observer performance • Lung, nodule, 68.332 • Lung neoplasms, diagnosis, 60.30
|
INTRODUCTION
|
|---|
The detection of nodular lesions on chest radiographs is an important task for radiologists. Because subtle nodules tend to be low in contrast and often overlap with ribs and clavicles, radiologists may fail to detect nodules in 20%–30% of actually positive cases (1,2). Kano et al (3) reported that the temporal subtraction technique, in which a previous image is subtracted from the current one, is useful for detecting a newly appearing lesion. This technique can cancel out normal bone, vascular, and mediastinal structures by using subtraction, and thus the detection accuracy of interval changes can be improved significantly (4). However, this technique is not applicable if a previous radiograph is not available.
We have developed a contralateral subtraction technique for enhancement of asymmetric opacities on chest radiographs by using subtraction of a reversed "mirror" image from the original. The contralateral subtraction image can cancel out symmetrical skeletal structures such as ribs and clavicles, except for the mediastinal and diaphragmatic regions, and thus can demonstrate a subtle nodule clearly without overlapping shadows. In this study, our purpose was to evaluate the potential usefulness of the contralateral subtraction technique for radiologists’ performance in the detection of subtle lung nodules on chest radiographs.
|
MATERIALS AND METHODS
|
|---|
Case Selection
From 247 chest images in the Japanese Standard Digital Image Database developed by the Japanese Society of Radiological Technology (5), two experienced chest radiologists (S.T., H.M.), who did not participate in the observer performance study, selected 25 chest images with subtle lung nodules and 25 normal chest images. In view of the availability of this database to researchers, institutional review board approval and informed consent were not required for this study. The absence and presence of nodules were confirmed by using CT findings. In this study, selection criteria for subtle lung nodules included low opacity, peripheral location; location apart from the mediastinum and diaphragmatic shadows; and location adjacent to or overlapping the bone shadows. Both the 25 lung nodule cases compatible with the selection criteria and the 25 normal cases were selected at random without the knowledge of the quality of subtraction images produced. The size of nodules ranged from 10 to 30 mm, with a mean of 16.8 mm. The final diagnosis of nodules included primary lung cancer (n = 17), metastatic lung tumor (n = 2), tuberculoma (n = 4), hamartoma (n = 1), and a noncharacterized nodule (n = 1). Chest images had been digitized with a 0.175-mm-pixel size, a matrix size of 2,048 x 2,048, and 12-bit gray-scale levels. However, in this study, the matrix size was reduced to 512 x 512 by subsampling of the original image data, and the number of gray-scale levels was decreased to 10 bits.
Computerized Scheme
The overall scheme of the contralateral subtraction technique is illustrated in Figure 1 (6). First, the lateral inclination of the chest image caused by improper patient positioning is corrected, because, if the midline of the thorax is inclined from the vertical direction, the difference in the angles of the midlines between the original and the reversed chest images will be doubled, which may lead to a serious misregistration error. To perform the lateral inclination correction, we applied an image rotation technique based on ribcage edge detection (4). The rotated image was laterally (right to left) flipped to produce a reversed "mirror" image. This mirror image was registered with the original image by using a nonlinear image-warping technique (3) to match the peripheral ribs in the mirror image with those in the original image. In this process, the mediastinal and cardiac regions are excluded because these structures are asymmetrical and because there is no useful information for image-matching and subtraction. Finally, the warped mirror image was subtracted from the original image, such as right ribs with left ribs and left ribs with right ribs, to produce the contralateral subtraction image. The algorithm was completely automated, and no user input was required. The computational time for producing a subtraction image with a personal computer was about 10 seconds. No cases were discarded in this study.
Display of Contralateral Subtraction Image
On the contralateral subtraction image, the subtracted lung is displayed by superimposition of the mediastinal and soft-tissue structures to resemble the original chest radiographs. It should be noted that a pulmonary nodule in the contralateral subtraction image is displayed as a dark area on the ipsilateral side and a light area on the contralateral side (Fig 2).
View larger version (146K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 2a. Comparison of chest image and corresponding contralateral subtraction image. (a) On the original chest radiograph, a nodule is noted in the left upper lung overlapping the distal clavicle. (b) On the contralateral subtraction image, symmetrical bone structures are canceled out, and then the nodule is clearly demonstrated without overlapping bone. The nodule is depicted as a dark area on the ipsilateral side (arrows) and as a bright area on the contralateral side. Asymmetrical structures such as the aortic arch and cardiac shadow are demonstrated as a dark area on the ipsilateral side and as white area on the contralateral side. Another dark area in the right upper lung field is an artifact due to misregistration.
|
|
View larger version (158K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 2b. Comparison of chest image and corresponding contralateral subtraction image. (a) On the original chest radiograph, a nodule is noted in the left upper lung overlapping the distal clavicle. (b) On the contralateral subtraction image, symmetrical bone structures are canceled out, and then the nodule is clearly demonstrated without overlapping bone. The nodule is depicted as a dark area on the ipsilateral side (arrows) and as a bright area on the contralateral side. Asymmetrical structures such as the aortic arch and cardiac shadow are demonstrated as a dark area on the ipsilateral side and as white area on the contralateral side. Another dark area in the right upper lung field is an artifact due to misregistration.
|
|
Observer Test Design
Twelve observers, including 10 attending radiologists (not specialized in chest radiology) and two residents, independently took part in an observer performance test. None of them had previous exposure to contralateral subtraction images, and, therefore, a training session was included prior to the test. We used a sequential test method in which each radiograph was shown first without the subtraction image for the initial rating. The subtraction image was then presented, and the observer was allowed to change his or her initial rating. Before the test, observers were informed concerning (a) the sequential test method to be used, (b) the role of the subtraction image as a "second opinion," (c) the nodule cases to be included in 50% of chest images, and (d) the definition of a nodule (ie, a focal, intrapulmonary, noncalcified soft-tissue opacity). In the observer test, chest radiographs were mounted on an alternator, and only one case at a time was shown. The reading time was not limited. A continuous rating scale that used a line-marking method was used to represent each observer’s confidence level by using a pencil on a line that was 7 cm in length. The left end of the line indicated complete confidence that the chest radiograph had no nodule, whereas the right end indicated complete confidence that the chest radiograph had a nodule. Intermediate levels of confidence were indicated by the position of a mark on the line, and locations close to the right and left ends indicated greater and lesser degrees of confidence, respectively, regarding the presence of a nodule. One author (S.K.) then measured the distance between the left end and the marked point and converted it to an ordinal confidence rating that ranged from 0 to 100.
In the sequential test, the initial rating without the subtraction image was marked with a blue pencil, and the second rating with the subtraction image was marked with a red pencil on the same line. Before the observer test, each observer underwent a training session that consisted of another six training cases, for familiarization with the observer test.
Data Analysis
The data were evaluated by using an algorithm (Metz LABROC4) (7), which calculates two parameters, a and b, representing (a) the maximum-likelihood estimates of the intercept and slope of the ROC curve when it is plotted on normal deviate axes and (b) the area under the ROC curve (Az). A receiver operating characteristic (ROC) curve representing a group of observers (entire group, residents, or attending radiologists) was then plotted by averaging of parameters a and b of each observer. The Az value ranges from zero to one and increases when the diagnostic performance approaches truth. The statistical significance of the difference between the Az value with and without the computer results was evaluated with a two-sided paired t test. The difference between the confidence ratings without and with the subtraction image was calculated for radiographs with and without nodules. Thus, for nodule cases, a positive difference implied a beneficial effect of the subtraction image, whereas a negative difference implied a detrimental effect of the subtraction image. For nonnodule cases, however, a positive difference implied a detrimental effect of the subtraction image, whereas a negative difference implied a beneficial effect of the subtraction image. We assumed that a clinically relevant change in the confidence rating occurred only when the difference was greater than 30 units on the 0–100 confidence rating scale.
|
RESULTS
|
|---|
Among the 12 observers enrolled in the study, one attending radiologist provided unusable data because a sufficiently wide range of confidence ratings was not used. This observer’s data were excluded, and thus a total of 11 observers remained in the study. The Table shows the Az values obtained both without and with subtraction images for each individual radiologist. The average ROC curves obtained are shown on Figures 3–5. It is apparent that the radiologists’ performance in nodule detection was improved with the contralateral subtraction images. The average Az value of all observers was improved significantly when the subtraction image was available (0.926–0.962, P = .011), except for two attending radiologists (observers B and G) who showed a decrease in Az value (from 0.987 to 0.986 and from 0.975 to 0.949, respectively). The average Az values for the two residents, obtained without and with subtraction images, were 0.827 and 0.926, respectively, whereas the average Az values for the nine attending radiologists, obtained without and with subtraction images, were 0.948 and 0.970, respectively. These differences were also statistically significant (P = .005 and P = .028, respectively), however, we recognize the small number of residents involved in this study. The gain in performance with the use of contralateral subtraction images was larger for residents than for attending radiologists. Figures 6 and 7 show the beneficial and detrimental effects of the contralateral subtraction images for each observer. The effect of the contralateral subtraction images was beneficial for all but one radiologist in improving their performance in nodule cases but was less effective in nonnodule cases. The average numbers of cases beneficially affected and detrimentally affected by using subtraction images in nodule cases were 1.6 and 0.5, respectively, whereas the average numbers of cases that were beneficially affected and detrimentally affected by using subtraction images in nonnodule cases were 2.3 and 2.1, respectively.
View larger version (30K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 3. Graph shows average ROC curves of all observers without (Az = 0.926) and with (Az = 0.962) subtraction images. Diagnostic performance is significantly improved (P = .012) with subtraction images.
|
|
View larger version (25K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 6. Graph shows number of cases affected by using subtraction images in nodule cases. Gray bars = beneficial effect, black bars = detrimental effect. Total numbers of cases beneficially affected and detrimentally affected by using subtraction images were 18 and five, respectively.
|
|
View larger version (21K):
[in this window]
[in a new window]
[Download PPT slide]
|
Figure 7. Graph shows numbers of cases affected by using subtraction images in nonnodule cases. Gray bars = beneficial effect, black bars = detrimental effect. Total numbers of cases beneficially affected and detrimentally affected by using subtraction images were 25 and 23, respectively.
|
|
|
DISCUSSION
|
|---|
The results of our observer test indicate that the contralateral subtraction technique can significantly improve radiologists’ performance in the detection of lung nodules on chest radiographs, especially for radiologists with limited experience. Before contralateral subtraction can be applied to clinical situations, some issues need to be addressed. In this subtraction technique, we excluded morphologic information on the mediastinum and diaphragm in creating the contralateral subtraction images because the mediastinum and diaphragm are asymmetrical structures and because there is no useful information for image matching and subtraction. It is important to note that this technique would be useful only for lesions in the periphery of the lungs away from the mediastinal and diaphragmatic regions. Another issue is an effective display method for lung nodules on contralateral subtraction images. On the contralateral subtraction image, a lung nodule is demonstrated as a dark area on the ipsilateral side and as a light area on the contralateral side. In our observer test, the observers viewed both lungs on the contralateral subtraction image, and we do not know which side of the lungs was actually used or which display mode was more useful for the detection of a nodule. Note that the information content of one lung field on the contralateral subtraction image is theoretically equivalent to that of the other, except that a nodule appears dark in one lung and light in the other.
In the beneficial analysis and detrimental analysis, the effect of the contralateral subtraction image was beneficial for all but one radiologists in improving their performance on nodule cases but was less effective in nonnodule cases. One possible cause for the less effectiveness in nonnodule cases might be artifacts in the normal lungs due to misregistration. Improper lateral inclination correction and failure in rib cage detection, image-matching, or the warping technique can produce misregistration artifacts. It is therefore necessary to improve the contralateral subtraction technique to reduce misregistration artifacts. The performance of one of the attending radiologists (radiologist G) decreased with use of the subtraction images (the Az without subtraction and with subtraction was 0.975 and 0.949, respectively). This result might have been due to the fact that radiologist G did not use the contralateral subtraction images well. Nevertheless, the average performance of the attending radiologists was improved significantly with the use of the subtraction images.
There is another type of subtraction technique, temporal subtraction, in which a previous image is subtracted from the current one, that is useful for detecting a newly appearing lesion. This technique also can cancel out normal bone, vascular, and mediastinal structures by using subtraction, and thus the detection accuracy of interval changes can be improved significantly. However, to perform this technique, a previous radiograph is required.
There are two methods for the ROC observer test, namely, an independent test and a sequential test. The independent test, which has been widely used, consists of two separate sessions and requires a longer time for completion. Kobayashi et al (8) reported that both methods, the independent test and the sequential test, provided the same conclusion and that there were no statistically significant differences in the Az values obtained with the two methods. Thus, in the current study, we used the sequential test, which can be performed in one session, without concern about the observer’s memory of chest images that were shown in a previous session.
In the future, a larger study with a variety of lung diseases should be conducted to examine the usefulness of the contralateral subtraction technique in clinical situations. In conclusion, our results indicate that the contralateral subtraction technique can significantly improve the diagnostic performance of radiologists in the detection of nodules on chest radiographs.
|
ACKNOWLEDGMENTS
|
|---|
We thank Yumiko Tsukiyama, MD, Kaiji Inoue, MD, Waka Saito, MD, Eito Kozawa, MD, Hisashi Hirata, MD, Junji Tanaka, MD, Makoto Amanuma, MD, Mihoko Yamazaki, MD, Masayuki Yuasa, MD, Naoko Nishi, MD, and Satomi Kawamoto, MD, for participating in observer tests and Katsumi Nakamura, MD, and Takayuki Ishida, PhD, for their helpful discussions and technical assistance.
|
FOOTNOTES
|
|---|
Abbreviation: ROC = receiver operating characteristic
Author contributions: Guarantors of integrity of entire study, S.T., K.D.; study concepts, S.T., S.K., K.D.; study design, S.T., K.D.; literature research, S.T.; clinical studies, S.T., A.H.; experimental studies, Q.L., K.D., S.K.; data acquisition, S.T., A.H.; data analysis/interpretation, S.T., S.K., K.D.; statistical analysis, S.T., S.K., K.D.; manuscript preparation, S.T., K.D.; manuscript definition of intellectual content, S.T., S.K., K.D.; manuscript editing, S.T., A.H., S.K., K.D.; manuscript revision/review and final version approval, all authors.
|
REFERENCES
|
|---|
-
Forrest JV, Friedman PJ. Radiological errors in patients with lung cancer. J Med 1981; 134:485-490.
-
Naidich DP, Zerhouni EA, Siegelman SS. Computed tomography of the thorax New York, NY: Raven, 1984; 171-199.
-
Kano A, Doi K, MacMahon H, Hassel DD, Giger M. Digital image subtraction of temporally sequential chest images for detection of interval change. Med Phys 1994; 21:453-461.[CrossRef][Medline]
-
Difazio MC, MacMahon H, Xu XW, et al. Digital chest radiology: effect of temporal subtraction images on detection accuracy. Radiology 1997; 202:447-452.[Abstract/Free Full Text]
-
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.[Abstract/Free Full Text]
-
Li Q, Katsuragawa S, Ishida T, et al. Contralateral subtraction: a novel technique for detection of asymmetric abnormalities on digital chest radiographs. Med Phys 2000; 27:47-55.[CrossRef][Medline]
-
Metz CE. Some practical issues of experimental design and data analysis in radiological ROC studies. Invest Radiol 1989; 24:234-245.[Medline]
-
Kobayashi T, Xu XW, MacMahon H, Metz CE, Doi K. Effect of a computer-aided diagnosis scheme on radiologists’ performance in detection of lung nodules on radiographs. Radiology 1996; 199:843-848.[Abstract/Free Full Text]
This article has been cited by other articles:
|
|
|
|
 
F. Li, R. Engelmann, K. Doi, and H. MacMahon
Improved Detection of Small Lung Cancers with Dual-Energy Subtraction Chest Radiography
Am. J. Roentgenol.,
April 1, 2008;
190(4):
886 - 891.
[Abstract]
[Full Text]
[PDF]
|
|
|
|
|
|
|
F. Li, R. Engelmann, C. E. Metz, K. Doi, and H. MacMahon
Lung Cancers Missed on Chest Radiographs: Results Obtained with a Commercial Computer-aided Detection Program
Radiology,
January 1, 2008;
246(1):
273 - 280.
[Abstract]
[Full Text]
[PDF]
|
|
|
|
|
|
|
M. B. Lobrano
Partnerships in Oncology and Radiology: The Role of Radiology in the Detection, Staging, and Follow-up of Lung Cancer
Oncologist,
July 1, 2006;
11(7):
774 - 779.
[Abstract]
[Full Text]
[PDF]
|
|
|
TOP
ABSTRACT
INTRODUCTION
