| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
Original Research |
1 Mallinckrodt Institute of Radiology, Washington University School of Medicine,
St. Louis, MO.
2 Present address: Department of Radiology, Seoul National University School of
Medicine, Bundang Hospital, Seoul, Korea.
3 Present address: Department of Radiology, University of Pittsburgh School of
Medicine, 200 Lothrop St., Suite 4895, Pittsburgh, PA 15213.
Received February 16, 2007;
accepted after revision June 7, 2007.
Address correspondence to K. T. Bae
(baek{at}upmc.edu).
Abstract
 Top Abstract IntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
MATERIALS AND METHODS. The cases of 30 patients with metastatic pulmonary nodules depicted on two serial clinical MDCT scans (16- or 64-MDCT, 5-mm section thickness) were studied. The number of nodules per patient varied from a minimum of two to innumerable. A maximum of 10 well-defined solid nodules per patient, a total of 210 nodules, were selected from each baseline CT scan and were evaluated for matching detection in follow-up CT by means of an automated program. Substantial changes in lung findings and lung volumes between serial scans were visually assessed. The effects on matching rate of interval lung changes and location, size, and total number of nodules in the lung were analyzed with contingency tables. Chi-square tests were used to evaluate patterns for statistical significance.
RESULTS. The nodule-matching rate per patient ranged from 0 to 100% (median, 87.5%). By nodule, the overall matching rate was 140 of 210 (66.7%). Matching rate was highly associated with changes in lung quality between serial studies. Matching of 122 of 148 nodules (82.4%) occurred in 23 patients with relatively unchanged lung findings, compared with 18 of 62 nodules (29.0%) in seven patients with substantial interval changes (p < 0.001). The matching rate decreased with an increased total number of nodules per lung. For 10 or fewer nodules per lung, matching was successful for 31 of 36 nodules; for 11–50 nodules per lung, 60 of 73 nodules; for 51–100 nodules per lung, 33 of 47 nodules; and for more than 100 nodules per lung, 16 of 54 nodules (p < 0.001). The matching rate was not significantly different with location or size of nodules.
CONCLUSION. The rate of automated matching of metastatic pulmonary nodules on clinical serial CT scans was high (82.4%) when the lung findings and lung expansion between the serial scans were relatively unchanged. The rate decreased significantly, however, with substantial interval changes in the lung and a larger number of nodules.
Keywords: chest CT • computer-aided diagnosis • oncology • pulmonary nodules
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
The lung is a frequent site of metastatic disease that manifests as pulmonary nodules. With sequential follow-up CT scans, changes in nodule size and number can be assessed [13, 14]. Evaluation of disease progression and treatment response, however, requires exact matching and precise quantitative analysis of nodules [15]. This task can be facilitated with automated matching and temporal assessment computer programs, which can improve the diagnostic performance and efficiency of radiologists.
Results of evaluations of software programs for the automated localization of pulmonary nodules on follow-up chest CT have been reported [14, 16, 17]. Although the studies emphasized the potential benefits of real-time automated matching for the follow-up of lung nodules on CT images, to our knowledge the performance of automated computed-aided matching has not been evaluated systematically for association with interval lung changes and nodule characteristics that may affect performance. The purpose of this study was to evaluate the performance of a computer-aided program used for automated matching of metastatic pulmonary nodules imaged in two serial clinical chest CT studies.
|
|
|
|
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
Patients
The study included 30 consecutively enrolled patients (11 men, 19 women;
age range, 24–80 years; mean, 57.4 years) with metastatic pulmonary
nodules depicted on two serial clinical chest CT studies. The primary sites of
malignancy were as follows: lung (n = 9), kidney (n = 4),
breast (n = 4), uterus (n = 4), prostate (n = 2),
colon (n = 1), rectum (n = 1), urinary bladder (n =
1), sphenoid sinus (n = 1), testicle (n = 1), mesentery
(n = 1), and unknown (n = 1). The mean interval between the
CT studies was 2.5 months (range, 10 days–5 months).
Pulmonary Nodule Selection
Only solid nodules larger than 3 mm and smaller than 2 cm in diameter were
included. No cavitary, ground-glass, or subsolid (i.e., partially solid and
partially ground-glass) nodules were present in the patient population. The
number of nodules per patient varied from a minimum of two to innumerable. Ten
patients had 10 or fewer nodules, nine patients had 11–50 nodules, five
patients had 51–100 nodules, and six patients had more than 100 nodules.
A maximum of 10 solid nodules per patient were sampled on baseline CT by a
radiologist with 10 years of experience in interpreting chest CT images,
resulting in a total of 236 nodules to review. We arbitrarily chose the
maximum of 10 nodules as a convenient sample size for inclusion of
representative nodules in each CT study. Nodules throughout the lungs were
selected arbitrarily but with consideration of representing nodules of varying
sizes and locations. Twenty-six of these nodules were excluded because on
follow-up CT they were either completely resolved or completely obscured by
surrounding infiltrates or atelectasis. The remaining 210 nodules were
included and evaluated for matched detection on follow-up CT performed with an
automated program.
CT
Routine contrast-enhanced clinical chest CT for metastasis evaluation was
performed with 120 mAs at 120 kVp and either a 16- or a 64-MDCT Sensation
scanner (Siemens Medical Solutions). The images were reconstructed with a
standard lung kernel at a section thickness of 5 mm with no intersection gap,
according to the routine clinical chest CT protocol at our institution.
Automated Matching of Pulmonary Nodules
Serial CT images were retrieved from the institutional PACS and sent to a
workstation (Leonardo, Siemens Medical Solutions) that contains an automated
matching program (LungCARE VB20, Siemens Medical Solutions)
[17,
18]. In brief, the
nodule-matching operation began with computation of approximate longitudinal
global alignment between the two serial sets of CT images. Refined alignment
parameters were then calculated on the basis of the cross-sectional area of
the lungs and the position of the trachea. Surface points of all surrounding
objects were extracted and used to produce a distance map. Points in the
follow-up set of CT images were superimposed onto the distance map for the
baseline set and then shifted in three directions in a search for the optimal
correlation between the two sets.
The performance of nodule matching with the program was assessed. When a nodule on the baseline CT images was marked by an operator with a mouse click, a volume of interest (VOI) surrounding the nodule was defined. At the same time, the corresponding VOI was determined automatically on the follow-up CT images (Fig. 1A, 1B, 1C, 1D). If the correctly matched nodule was located within the VOI on the follow-up CT images, automated matching was considered successful. In addition to the transverse images, coronal reformatted images were displayed as a visual aid to assessment of automated nodule matching.
Data and Statistical Analysis
Substantial interval changes in the lung findings (effusion, atelectasis,
infiltrates, and marked changes in disease state) or lung expansion between
serial CT studies were visually assessed by consensus of two radiologists with
10 and 11 years of experience in interpreting chest CT images. The patients
were divided into two groups, those with relatively unchanged and those with
substantially changed lung findings. Substantial changes included considerable
interval differences in lung expansion and the extent of lung cancer and
metastatic masses, lung infiltrates, pleural effusion, and atelectasis. The
rate of nodule matching was evaluated, and the rates for the two groups were
compared.
The patients were divided into four groups according to the total number of nodules within their lungs: 10 or fewer nodules, 11–50 nodules, 51–100 nodules, and more than 100 nodules. The nodule-matching rates of the four groups were compared. Each nodule was characterized according to its location within the lung as parenchymal, peripheral, juxtaphrenic, or juxtavascular. Parenchymal nodules were completely surrounded by aerated lung parenchyma. Peripheral nodules were contacting or located within 2 mm of the pleura. Juxtaphrenic nodules were in contact with the diaphragm. Juxtavascular nodules were in contact with a blood vessel. The nodule-matching rates for these four nodule locations were compared. The nodules were divided into two groups according to size on baseline CT. The size (maximum diameter) of a nodule was computed automatically from its segmented nodule volume with the matching program. We arbitrarily used 10 mm as the size cutoff for the two groups: nodules 10 mm or less and nodules greater than 10 mm in maximum diameter. The nodule-matching rates of the two groups were compared.
We constructed contingency tables to compare differences in ability to match nodules by patient and nodule characteristics. Patterns were tested for statistical significance with chi-square tests. We used the program SPSS version 12.0 for Windows (SPSS) for the statistical analysis.
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
We observed that the lung findings were relatively unchanged in 23 patients but that substantial changes had occurred in the other seven patients: markedly changed expansion of the lung (n = 1) (Fig. 2A, 2B, 2C, 2D), markedly changed extent of lung cancer or metastatic masses (n =4) (Fig. 3A, 3B, 3C, 3D), and markedly increased amount of pleural effusion with atelectasis (n =2). The matching rate was highly associated with the changes in lung quality between serial scans (p < 0.001): 122 of 148 nodules (82.4%) were matched in 23 patients with relatively unchanged lung findings, compared with 18 of 62 nodules (29.0%) in seven patients with substantial interval changes (Table 1).
|
|
|
|
|
|
|
|
|
The nodule-matching rate tended to decrease slowly with increasing numbers of nodules but decreased dramatically when the number of nodules was greater than 100 (p < 0.001 for the differences between the four nodule-number groups). The matching rate was 31 of 36 (86.1%) for the group with 10 or fewer nodules, 60 of 73 (82.2%) for the group with 11–50 nodules, 33 of 47 (70.2%) for the group with 51–100 nodules, and 16 of 54 (29.6%) for the group with more than 100 nodules (Table 2).
|
The matching rates did not differ significantly with nodule location (p = 0.87): 67.2% (88/131) for the parenchymal nodules, 62.5% (25/40) for the peripheral nodules, 64.7% (11/17) for the juxtaphrenic nodules, and 72.7% (16/22) for the juxtavascular nodules. The matching rate also was not significantly different between the large-nodule and small-nodule groups (p = 0.27): 69.6% (87/125) for nodules 10 mm or smaller and 62.4% (53/85) for nodules larger than 10 mm.
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
Results of several clinical evaluations of software for automated localization of lung nodules on follow-up CT examinations have been published [14, 16, 17]. The most recent study [17] showed that the automatic matching rate was 86.3% and that the matching rate was not affected by the location or diameter of nodules. Compared with that rate, the overall automatic matching rate in our study was lower, 66.7%. This discrepancy in the matching rates is likely related to our study population, which included patients with marked disease progression and with substantially variable inspiration levels. This postulation may be supported by the fact that the nodule-matching rate among patients with relatively unchanged disease extent and inspiration level was 82.4%, comparable with the rate in the previous study [17]. Our study differed from the previous study in that we performed an additional evaluation and found that the matching rate was significantly affected by interval changes in disease and total number of metastatic nodules within the lungs.
Matching rate was highly associated with changes in lung quality between serial CT studies. The matching rate was only 18 of 62 nodules (29%) in seven patients with substantial interval changes. As a result, use of this automatic matching algorithm should be confined to follow-up images with similar conditions of surrounding lung disease and similar inspiratory states. This algorithm would be expected to perform well in the population undergoing CT for lung cancer screening because lung image quality and findings usually remain stable between serial CT studies. The algorithm may have to be improved further to achieve a high degree of matching in serial CT studies in which large masses, marked interval changes, and innumerable metastatic nodules are found.
In general, the sensitivity for nodule detection with CAD diminishes with decreasing nodule size [11, 19]. The performance of CAD systems in nodule detection also depends on the anatomic location of pulmonary nodules [12]. In particular, parenchymal and isolated nodules are more reliably detected than juxtapleural and juxtavascular nodules, because parenchymal nodules are surrounded by distinctly lower-attenuation pulmonary parenchyma [12]. In contrast, the performance of nodule matching with CAD is not influenced significantly by the location or size of nodules, as found in our study. This discrepancy in the performances of nodule detection and matching programs is not surprising. Although the nodule detection tasks mainly involve detection and differentiation of nodules from surrounding vessels, pleura, and airways, nodule-matching tasks mainly depend on the search for and determination of similar nodule structures within a predefined search space from a selected nodule position. Therefore, some geometric features of nodules, such as size, shape, and location, that are imperative for detection of nodules may be less crucial in nodule matching.
In this study, the matching rate tended to decrease slowly with increasing numbers of nodules but decreased dramatically when there were numerous nodules in the lung. The low matching rate with a large number of nodules was caused by the presence of unmatched nodules rather than by incorrect matching of nodules. Although we are uncertain how the number of nodules within the lungs affects the operation of the matching algorithm, our finding suggests that registration of the coordinates between two serial CT scans is disturbed, likely because many nodules occupy the lung parenchyma and obscure the normal anatomic landmarks needed for registration. Another less likely possibility is that the accuracy of the correspondence between two CT studies is reduced owing to the presence of multiple matching candidates within the search VOI.
For clinical applications, it appears that the current automated matching program would allow faster assessment of whether metastatic disease has changed than would assessment by manual matching of nodules. Although the matching rates are reduced in some circumstances, the program seems to match enough nodules to be of benefit in the sense that it can reduce the radiologist's time spent determining whether change has occurred. Most of the unmatched nodules were found within three sections of the nodule to be compared, potentially reducing the search time even for cases in which a nodule selected for follow-up comparison is not perfectly matched. Further study is required for assessment of the clinical benefits of automated nodule matching.
This study had limitations. First, it was difficult to objectively define interval changes in lung quality. We relied on the consensus interpretation of two chest radiologists for evaluation of lung quality, including any differences in inspiratory levels between serial CT examinations. Second, only well-defined solid nodules, up to 10 nodules per patient, were arbitrarily selected and included in our study. No ground-glass opacity or alveolar infiltrative or cavitary nodules were included. The detection and matching of these nonsolid nodules likely would be more complicated, and further studies should be pursued. Third, the clinical importance of nodule mismatching was not assessed. This study was a retrospective evaluation of the performance of an automated lung-matching program. We either identified mismatched nodules by carefully comparing and searching for nodules between serial CT studies or detected lung findings that caused nodule mismatching. This study was not designed to compare the performances of the CAD program and the radiologists.
In conclusion, the rate of automated matching of metastatic pulmonary nodules in clinical serial CT scans was high (82.4%) when the lung findings and expansion between the serial scans were relatively unchanged. The rate decreased significantly, however, with substantial interval lung changes and an increased number of nodules. Thus, in the current implementation, application of the automatic matching algorithm seems most appropriate for follow-up images obtained in cases with similar conditions of surrounding lung disease and similar inspiratory states. The radiologist's attention to the accuracy of the matching result for each nodule is still recommended in the clinical setting. The importance of automated nodule matching to clinical care and workflow remains to be determined.
Acknowledgments
We thank Thomas K. Pilgram for reviewing and commenting on the statistical
analysis method and results of our study.
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
CiteULike 
Complore 
Connotea 
Del.icio.us 
Digg 
Reddit 
Technorati What's this?
This article has been cited by other articles:
|
|
 |
|
  C. Tao, D. S. Gierada, F. Zhu, T. K. Pilgram, J. H. Wang, and K. T. Bae Automated Matching of Pulmonary Nodules: Evaluation in Serial Screening Chest CT Am. J. Roentgenol., March 1, 2009; 192(3): 624 - 628. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
