HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS		QUICK SEARCH:   [advanced] Author: Keyword(s): Year:  Vol:  Page:

This Article Abstract Figures Only Full Text (PDF) All Versions of this Article: 2361041286v1 236/1/286    most recent Submit a response Alert me when this article is cited Alert me when eLetters are posted Alert me if a correction is posted Citation Map Services Email this article to a friend Similar articles in this journal Similar articles in PubMed Alert me to new issues of the journal Download to citation manager Citing Articles Citing Articles via HighWire Citing Articles via Google Scholar Google Scholar Articles by Bae, K. T. Articles by Kim, J.-H. Search for Related Content PubMed PubMed Citation Articles by Bae, K. T. Articles by Kim, J.-H.

Pulmonary Nodules: Automated Detection on CT Images with Morphologic Matching Algorithm—Preliminary Results¹

¹ From the Mallinckrodt Institute of Radiology, Washington University School of Medicine, 510 S Kingshighway Blvd, St Louis, MO 63110. Received July 23, 2004; revision requested September 28; revision received October 23; accepted December 10. Address correspondence to K.T.B. (e-mail: baet{at}mir.wustl.edu).

	ABSTRACT

TOP ABSTRACT INTRODUCTION Materials and Methods Results Discussion References

 
Institutional review board approval was obtained. Informed patient consent was not required for this retrospective study, which involved review of previously obtained image data. Patient confidentiality was protected; the study was compliant with the Health Insurance Portability and Accountability Act. An automated pulmonary nodule detection program that takes advantage of three-dimensional volumetric data was developed and tested with multi–detector row computed tomographic (CT) images from 20 patients (13 men, seven women; age range, 40–75 years) with pulmonary nodules. A total of 164 nodules 3 mm in diameter and larger were detected by two radiologists in consensus and were used as a reference standard to evaluate the computer-aided detection (CAD) program. The CAD algorithm was structured to process nodules that were categorized into three types: isolated, juxtapleural, and juxtavascular. Overall sensitivity for nodule detection with the CAD program was 95.1% (156 of 164 nodules). The sensitivity according to nodule size was 91.2% (52 of 57 nodules) for nodules 3 mm to less than 5 mm and 97.2% (104 of 107 nodules) for nodules 5 mm and larger. The number of false-positive detections per patient was 6.9 for false nodule structures 3 mm and larger and 4.0 for false nodule structures 5 mm and larger.

© RSNA, 2005

	INTRODUCTION

TOP ABSTRACT INTRODUCTION Materials and Methods Results Discussion References

 
Lung cancer is the second most commonly diagnosed cancer in the United States, and the lung is a frequent site of metastasis from other cancers that manifest as pulmonary nodules. Chest computed tomography (CT) is the most sensitive diagnostic imaging modality for the detection of lung cancer and the resolution of any equivocal abnormalities detected on chest radiographs (1,2). With spiral CT and, more recently, multi–detector row CT techniques, the sensitivity for detection of pulmonary nodules has improved further (3,4). Recently, CT techniques have been applied to screening for lung cancer in high-risk populations and have been shown to be promising for detection of early lung cancers (5). Thin-section three-dimensional (3D) CT of the thorax may allow us to evaluate small nodules at early stages. With sequential follow-up CT scans, early changes in nodule size and number can be assessed (6,7).

A thin-section CT scan of the whole thorax generates a large data set, typically 250–350 images of 1-mm section thickness, and requires radiologists to spend a considerable amount of time interpreting the images. As a means to reduce radiologists' workload, computer-aided detection (CAD) systems may be used. The CAD systems can also help improve a radiologist's performance in the detection of pulmonary nodules (8).

Various CAD systems for pulmonary nodule detection at CT have been proposed since the report of a system by Bae et al (9) more than a decade ago. The nodule detection algorithms of earlier CAD systems (7,10–15) were based on two-dimensional (2D) morphologic features on each section, serial 2D methods used to examine connectivity of the features on adjacent sections, and/or limited 3D structural information.

More recent CAD systems for the detection of nodules (16,17) have been reported to use 3D algorithms that take advantage of thin-section volumetric CT data. We postulate that an efficient CAD algorithm can be developed by combining both 3D and 2D morphologic operations in the segmentation, detection, and classification of pulmonary nodules. Thus, the objective of this study was to develop and test a CAD program based on combined 3D and 2D morphologic operation algorithms that would allow us to detect pulmonary nodules, with full automation and high efficiency, based on CT volumetric data acquired from multi–detector row CT.

	Materials and Methods

TOP ABSTRACT INTRODUCTION Materials and Methods Results Discussion References

 
Institutional review board approval was obtained. Informed patient consent was not required for our retrospective study, which involved only review of previously obtained image data. Patient confidentiality was protected; our study was compliant with the Health Insurance Portability and Accountability Act.

Patients and CT Image Data
Our database consisted of diagnostic thoracic CT scans from 20 consecutive patients with pulmonary nodules (13 men, seven women; age range, 40–75 years; mean age, 57 years) who were referred to undergo thin-section CT scanning for the evaluation of pulmonary nodules (Fig 1). In retrospectively selecting the study patients, we did not target any specific diagnosis or pathologic condition that was manifested by pulmonary nodules. The nodules could be associated with metastasis, primary lung cancer, or granulomas. Both calcified and noncalcified nodules were included. No cavitary, ground-glass, or subsolid nodules were present in our patient population. CT studies with substantial parenchymal or pleural diseases such as consolidation, fibrosis, and pleural effusion were excluded from our study population.

View larger version (26K): [in this window] [in a new window] [Download PPT slide]   Figure 1. Bar graph shows the number and size distribution of lung nodules in 20 patients. A total of 164 nodules was identified. Eighteen of 20 patients had 1–13 pulmonary nodules (mean, 4.7 nodules), and the other two patients had 25 and 54 nodules on CT images.

CAD Program
Overview and organization of the CAD program.—The lung nodule detection algorithm is composed of a combination of 2D and 3D operations. The 3D operations are more general and powerful, but in some image processing steps, the 2D operations are preferred because they are more efficient and easier to implement. This combined 2D and 3D approach may be used by radiologists when the CT images are reviewed for the detection of lung nodules. Isolated nodules, which are nodules that are dissimilar to adjacent blood vessels and have good tissue contrast from the surrounding much-lower-attenuation pulmonary parenchyma, can be identified easily on individual 2D sections without the need for comparison of adjacent sections. However, juxtavascular nodules—nodules that are closely associated with or attached to adjacent blood vessels—may not be detected easily because some blood vessels perpendicular to the imaging plane may appear as nodules on 2D sections. Comparison with adjacent sections—that is, 3D operation—is essential for the detection of these lung nodules. Juxtapleural nodules, which are pulmonary nodules attached to the pleural surface, may be detected easily on 2D sections by means of human perception, but an additional operation is required in CAD to separate them from the pleural boundary. Our CAD algorithm was designed and organized to process different types of nodules by applying sets of 2D and 3D operations unique to each nodule type (Fig 2).

View larger version (36K): [in this window] [in a new window] [Download PPT slide]   Figure 2. Flow diagram illustrates the overall scheme for automated lung nodule detection on CT images.

Thorax and lung region segmentation.—The nodule detection system begins by automatically segmenting the thorax and lung region from the CT images. The thorax can be delineated easily from background air because of its high tissue-air contrast. For the same reason, the lung can be distinctly separated from surrounding soft tissues or bone. Gray-level thresholding methods have commonly been used to create binary images and then detect the boundary of the thorax and lung regions (7,10,13,18). The threshold can be determined from a fixed attenuation value, an analysis of the image attenuation profile, or an analysis of a histogram of a 2D or 3D region. Our segmentation for the thorax and lung region was also based on a thresholding method in each section. The threshold was determined by analyzing the 2D region histogram, which showed distinct groups of pixels belonging to the thorax and background air. The thoracic region was segmented from the thresholded binary image. Background pixels representing air outside the thorax were clipped, and the morphologic operations of erosion and dilation were applied to eliminate scattered background. Then, a new gray-level histogram was acquired from the pixels in the entire thoracic region, and the threshold value for lung segmentation was estimated to maximize the separation between two major peaks on the histogram. The boundary detection processes were repeated in the segmented thorax region to segment the lung region.

Lung boundary revision.—The lung boundary detected with the gray-level thresholding technique often does not include juxtapleural nodules because these nodules are contiguous with the body wall and are thus segmented as a part of the body wall instead of the lung region. In addition, central pulmonary vessels with any attached juxtavascular nodules may be excluded from the segmented lung region. A morphologic operation (13,19) was performed to correct such segmentation "errors" in the lung boundary. Circular "closing" filters of variable diameters were applied iteratively to the segmented lung region to capture the juxtapleural nodules and to include central pulmonary vessels. The closing filter consists of a dilation followed by erosion morphologic operations and is used to fill in holes and small gaps (20). The lung region segmentation process, including lung boundary revision, is illustrated in Figure 3. The final segmented lung region contains three types (isolated, juxtapleural, and juxtavascular) of lung nodules, blood vessels, pulmonary parenchyma, airway, and some mediastinal structures.

View larger version (149K): [in this window] [in a new window] [Download PPT slide]   Figure 3a. Patient 5. Intermediate 2D CT images illustrate the lung region segmentation process. (a) Transverse CT image of the thorax underwent gray-level thresholding to generate (b) a binary image. (c) An initial uncorrected lung region was estimated. (d) After a corrected lung boundary was determined, (e) the initial CT image enclosed by this boundary was isolated. (f) The soft-tissue structures within the lung region were segmented by means of gray-level thresholding.

View larger version (52K): [in this window] [in a new window] [Download PPT slide]   Figure 3b. Patient 5. Intermediate 2D CT images illustrate the lung region segmentation process. (a) Transverse CT image of the thorax underwent gray-level thresholding to generate (b) a binary image. (c) An initial uncorrected lung region was estimated. (d) After a corrected lung boundary was determined, (e) the initial CT image enclosed by this boundary was isolated. (f) The soft-tissue structures within the lung region were segmented by means of gray-level thresholding.

View larger version (42K): [in this window] [in a new window] [Download PPT slide]   Figure 3c. Patient 5. Intermediate 2D CT images illustrate the lung region segmentation process. (a) Transverse CT image of the thorax underwent gray-level thresholding to generate (b) a binary image. (c) An initial uncorrected lung region was estimated. (d) After a corrected lung boundary was determined, (e) the initial CT image enclosed by this boundary was isolated. (f) The soft-tissue structures within the lung region were segmented by means of gray-level thresholding.

View larger version (32K): [in this window] [in a new window] [Download PPT slide]   Figure 3d. Patient 5. Intermediate 2D CT images illustrate the lung region segmentation process. (a) Transverse CT image of the thorax underwent gray-level thresholding to generate (b) a binary image. (c) An initial uncorrected lung region was estimated. (d) After a corrected lung boundary was determined, (e) the initial CT image enclosed by this boundary was isolated. (f) The soft-tissue structures within the lung region were segmented by means of gray-level thresholding.

View larger version (110K): [in this window] [in a new window] [Download PPT slide]   Figure 3e. Patient 5. Intermediate 2D CT images illustrate the lung region segmentation process. (a) Transverse CT image of the thorax underwent gray-level thresholding to generate (b) a binary image. (c) An initial uncorrected lung region was estimated. (d) After a corrected lung boundary was determined, (e) the initial CT image enclosed by this boundary was isolated. (f) The soft-tissue structures within the lung region were segmented by means of gray-level thresholding.

View larger version (44K): [in this window] [in a new window] [Download PPT slide]   Figure 3f. Patient 5. Intermediate 2D CT images illustrate the lung region segmentation process. (a) Transverse CT image of the thorax underwent gray-level thresholding to generate (b) a binary image. (c) An initial uncorrected lung region was estimated. (d) After a corrected lung boundary was determined, (e) the initial CT image enclosed by this boundary was isolated. (f) The soft-tissue structures within the lung region were segmented by means of gray-level thresholding.

View larger version (122K): [in this window] [in a new window] [Download PPT slide]   Figure 4a. Patient 15. Segmented 3D volumetric data (a) before and (b) after removing the nonvessel group. Image a was obtained by applying region growing and labeling to a stack of 2D segmented lung images (one of which is shown in Fig 3f). Image a contains three types of lung nodules (isolated, juxtapleural, and juxtavascular), blood vessels, and noise voxels. The 3D data set in b—that is, the vessel group, including juxtavascular nodules—represents a subset of the 3D data in a, after the structures not connected to the pulmonary vessels were removed.

View larger version (95K): [in this window] [in a new window] [Download PPT slide]   Figure 4b. Patient 15. Segmented 3D volumetric data (a) before and (b) after removing the nonvessel group. Image a was obtained by applying region growing and labeling to a stack of 2D segmented lung images (one of which is shown in Fig 3f). Image a contains three types of lung nodules (isolated, juxtapleural, and juxtavascular), blood vessels, and noise voxels. The 3D data set in b—that is, the vessel group, including juxtavascular nodules—represents a subset of the 3D data in a, after the structures not connected to the pulmonary vessels were removed.

Nodules are typically compact and spherical. The compactness of each nodule candidate was computed as the ratio of its volume to the volume of the smallest 3D box that encloses the candidate structure. A candidate was considered to be a nonnodule if its compactness value was less than 0.5 or greater than 1.5. An elongation factor was computed for each candidate in both 2D and 3D. The 2D elongation factor corresponded to the distance ratio of the major axis to the minor axis of a rectangle or ellipse enclosing the candidate. The 3D elongation factor was computed by using the ratio of the maximum to the minimum eigenvalues that were calculated from the voxel coordinates of the candidate structure. The elongation factor becomes close to 1 for a relatively round or spherical structure. In the current version of our implementation, a candidate with an elongation factor of greater than three was considered to be a nonnodule. The remaining nodule candidates that were not excluded according to size, compactness, and elongation criteria from the nonvessel group were subsequently classified as nodules.

Detection of juxtavascular nodules.—Juxtavascular nodules are distinguished morphologically from the vessels because they are typically spherical in shape, while the vessels are elongated. Detection of juxtavascular nodules, however, is more difficult than that of isolated or juxtapleural nodules because juxtavascular nodules are attached to the pulmonary vessels. Shape and geometric features of juxtavascular nodules cannot be characterized independently from the vessels without disassociating them from the attached vessels. Direct application of shape criteria, such as compactness or elongation factor, is ineffective for the identification of nodule candidates.

We used a 3D multilevel morphologic matching method to identify and extract juxtavascular nodule candidates from the vessel group. The 3D morphologic filters were configured. These filters were spherical in shape, and four different sizes of the filters (3, 6, 9, and 12 mm in diameter) were used to identify nodule candidates ranging from 3 to 30 mm. Although the 3-mm filter was the most sensitive among the filters, and it detected many nodule candidates larger than 3 mm, use of the other filters was essential to identify nodule candidates that were subject to high variability in pulmonary vessel thickness in different lung regions. Each filter was particularly sensitive in the detection of nodule candidates whose diameters were close to that of the corresponding filter. Increasing the number of filters improved the sensitivity of nodule candidate detection but increased computation time and the number of false-positive candidates detected.

The morphologic filters were convolved with the 3D data of the vessel group to compute correlation (19,22). The correlation between 3D data and morphologic filter can be expressed in terms of the fast Fourier transform method (23). After the correlation was computed, juxtavascular nodule candidates were selected by performing thresholding of the correlation values. The threshold value was determined empirically at 70%. The use of lower threshold values, as with the use of an increasing number of morphologic filters, increased the sensitivity of detecting nodule candidates but at the price of increasing computation time and the number of false-positive candidates. After the multilevel matching process, 3D region growing and labeling were applied to the nodule candidates. The 3D shape and geometric features were computed for each candidate. Candidates classified as nonnodules were eliminated by means of the rule-based scheme, as described in the previous section, and the remaining candidates were flagged as nodules.

Image Interpretation
The CT images were initially reviewed on a clinical workstation by one thoracic radiologist (J.H.K.), who has 14 years of experience in interpreting chest CT images. The number, size, and location of nodules were recorded without the use of the CAD program. Three to 4 weeks after this initial reading, two radiologists (J.H.K., K.T.B.), who have 14 and 12 years of experience in interpreting chest CT images, respectively, established the reference standard by means of consensus. The 3–4 week interval between the readings of any given image was implemented to reduce the possibility of the memory of the first reading affecting the second reading. The cases were presented in a different order from that in the first reading. Each reviewer first searched freely through the CT images that included the findings of pulmonary nodules detected by the CAD program and performed the detection and documentation of pulmonary nodules. Then, the reviewers assessed together each detected finding on the CT images and arrived at a final decision by means of consensus if the finding was a true nodule or a false-positive finding.

The number, size, and location of nodules were documented. The size of the nodules was measured by using the electronic caliper in the image display software. A total of 164 consensus nodules were identified. Eighteen of 20 patients had 1–13 pulmonary nodules (mean, 4.7 nodules) on CT images, and the other two patients had 25 and 54 nodules. Of the 164 nodules, 27 were 10 mm and larger, 80 were 5 mm to less than 10 mm, and 57 were 3 mm to less than 5 mm. The largest nodule was 27 mm. Eight of 164 nodules were calcified. A detailed distribution of the number and size of the pulmonary nodules in each patient is shown in Figure 1.

Data and Statistical Analysis
The number, size, and location of nodules detected by the CAD program alone and by means of the radiologists' consensus reading of images with the CAD system were tabulated (J.S.K.). The detected nodules were divided into three groups according to diameter (3 mm to less than 5 mm, 5 mm to less than 10 mm, 10 mm and larger) and location (isolated, juxtapleural, juxtavascular). By using the radiologists' consensus reading augmented with CAD as the reference standard, the sensitivity and the number of false-positive findings detected with the CAD system alone were computed for each group and for the sum of the groups. In addition, the initial radiologist's reading was compared with the detection by the CAD program alone and by the reference standard. Statistical analyses were performed with JMP Statistical Software, version 5.1 (SAS; Cary, NC).

	Results

TOP ABSTRACT INTRODUCTION Materials and Methods Results Discussion References

 
The overall sensitivity for nodule detection (Figs 5, 6) with our method was 95.1% (156 of 164 nodules) for all nodules 3 mm and larger (Table). The sensitivity for detecting nodules according to category was 97.4% (76 of 78 nodules) for isolated nodules, 92.3% (48 of 52 nodules) for juxtapleural nodules, and 94.1% (32 of 34 nodules) for juxtavascular nodules. The sensitivity for detecting nodules according to size was 97.2% (104 of 107 nodules) for nodules 5 mm and larger and 91.2% (52 of 57 nodules) for nodules 3 mm to less than 5 mm. Nodules that were not detected with CAD—that is, false-negative nodules—were either juxtapleural nodules attached obtusely along the pleura (n = 3) or juxtavascular nodules (n = 2) that were incompletely segmented or weakly visible nodules (n = 3) due to low contrast resolution or volume averaging.

View larger version (146K): [in this window] [in a new window] [Download PPT slide]   Figure 5a. Patient 5. (a, c) Transverse CT images (b, d) with nodules detected by the CAD system. Image b shows juxtavascular (arrow) and isolated (arrowhead) nodules detected on the CT image in a. Image d shows two juxtapleural nodules (arrows) detected on the CT image in c. No false-negative or false-positive findings are observed on these two CT images.

View larger version (144K): [in this window] [in a new window] [Download PPT slide]   Figure 5b. Patient 5. (a, c) Transverse CT images (b, d) with nodules detected by the CAD system. Image b shows juxtavascular (arrow) and isolated (arrowhead) nodules detected on the CT image in a. Image d shows two juxtapleural nodules (arrows) detected on the CT image in c. No false-negative or false-positive findings are observed on these two CT images.

View larger version (149K): [in this window] [in a new window] [Download PPT slide]   Figure 5c. Patient 5. (a, c) Transverse CT images (b, d) with nodules detected by the CAD system. Image b shows juxtavascular (arrow) and isolated (arrowhead) nodules detected on the CT image in a. Image d shows two juxtapleural nodules (arrows) detected on the CT image in c. No false-negative or false-positive findings are observed on these two CT images.

View larger version (145K): [in this window] [in a new window] [Download PPT slide]   Figure 5d. Patient 5. (a, c) Transverse CT images (b, d) with nodules detected by the CAD system. Image b shows juxtavascular (arrow) and isolated (arrowhead) nodules detected on the CT image in a. Image d shows two juxtapleural nodules (arrows) detected on the CT image in c. No false-negative or false-positive findings are observed on these two CT images.

View larger version (113K): [in this window] [in a new window] [Download PPT slide]   Figure 6a. Patient 15. (a) Anterior and (b) posterior views of a 3D volumetric representation of pulmonary vessels and detected lung nodules (arrows). This patient had a total of 13 nodules (seven isolated, three juxtapleural, and three juxtavascular nodules), all of which were detected with the CAD system. There were two false-positive findings in this case.

View larger version (124K): [in this window] [in a new window] [Download PPT slide]   Figure 6b. Patient 15. (a) Anterior and (b) posterior views of a 3D volumetric representation of pulmonary vessels and detected lung nodules (arrows). This patient had a total of 13 nodules (seven isolated, three juxtapleural, and three juxtavascular nodules), all of which were detected with the CAD system. There were two false-positive findings in this case.

View larger version (33K): [in this window] [in a new window] [Download PPT slide]   Figure 7. Bar graph shows the number of lung nodules detected by the CAD system and a chest radiologist (first reading without CAD and second reading with CAD) in 20 patients. Lung nodule detection by the CAD system was highly accurate and better than the radiologist's first reading without CAD.

	Discussion

TOP ABSTRACT INTRODUCTION Materials and Methods Results Discussion References

In the past decade, various CAD systems have been proposed for the detection of pulmonary nodules on CT images (7,10,11,13–17). The majority of CAD systems were developed and evaluated on single–detector row CT images with 5–10-mm section thickness (7,10,11,13–15), while more recent systems were assessed with thin-section single– (17) or multi–detector row CT images (16). The sensitivity for detecting nodules with reported CAD systems varied from 38% to 100%, and the false-positive detections per case ranged from 1 to 75. The performance of CAD systems appears to be highly associated with the section thickness (or reconstruction interval) and seems to be better on thin-section than on thick-section CT images (27).

In general, the sensitivity for nodule detection decreased with decreasing nodule size (7,16,17). Brown et al (17) reported 70% sensitivity for the detection of micronodules (1–3 mm) in contrast to 100% sensitivity for nodules larger than 3 mm. The sensitivity of the CAD system used by Ko and Betke (7) decreased from 91% for nodules larger than 3 mm to 86% for all nodules, including those smaller than 3 mm. A similar trend, shown by Qian et al (16), was that the sensitivity decreased from 87.1% for solid nodules 3 mm and larger to 77.1% for all nodules, including ones smaller than 3 mm. Likewise, in our study, where only the nodules 3 mm and larger were considered, the sensitivity decreased with the size of nodules: from 97.2% for nodules 5 mm and larger to 95.1% for all nodules 3 mm and larger. Furthermore, lowering the size criteria for nodules increases the number of false-positive findings, in addition to decreasing the sensitivity, which results in an overall reduction in CAD performance. It is noteworthy, however, that the performance of radiologists in the detection of micronodules (1–3 mm) was markedly lower than in the detection of larger nodules. Also, the relative contribution of CAD to improving the radiologists' performance for detection of small nodules was substantially higher than that for detection of larger nodules (16,17).

The performance of CAD also depends on the anatomic location of pulmonary nodules. The performance breakdown according to nodule location, however, was rarely reported in previous CAD studies. In our study, we recognize that pulmonary nodules can be divided into three types in terms of their proximity to surrounding anatomic structures: isolated, juxtapleural, and juxtavascular nodules. Detection of isolated nodules is easier than that of the other two types in terms of both human perception and computer detection, because the isolated nodules are surrounded by distinctly lower-attenuation pulmonary parenchyma. In contrast, detection of juxtapleural and juxtavascular nodules is more challenging because these nodules may not be easily differentiated from abutting soft-tissue structures. Each of the three types of nodules requires a unique algorithmic approach for detection. This consideration is critical for implementing a CAD system.

A variety of algorithmic approaches has been proposed to develop CAD systems for the detection of pulmonary nodules (7,10,11,13–17,27,28). In many CAD systems, segmentation of lung regions from the thorax constitutes the first step of image processing. Some researchers (17,27) partitioned the segmented lung regions further into central and peripheral subregions. Juxtapleural nodules were often excluded from the initial extracted lung regions and would require an additional segmentation step, such as the morphologic operation that we used in our study. Other reported techniques for this segmentation process include the rolling ball algorithm (13), lung boundary curvature analysis (7,11), boundary distance analysis (27), and template matching (14,16). In our study, juxtavascular nodules were dissociated from the vessels by means of a 3D morphologic matching and correlation method. Previously reported approaches for the same application include multiple gray-level thresholding (7,10,13), local histogram analysis (16), and a morphologic opening operation (17).

After nodule candidates were identified, nodules were classified from nonnodule structures on the basis of the features of the candidates. Rule-based decision approaches such as ours were used most commonly (7,10,13,14,16,27), while Brown et al (17) used a model-based, fuzzy logic approach to classify nodules. Rule-base schemes in which the classification process is based on thresholds for features of nodule candidates are widely used because they are more straightforward to implement than other methods, such as those based on statistical methods, neural networks, and fuzzy logic (29).

In our study and others, threshold values and criteria were usually determined heuristically and empirically from spherical nodule geometric constraints, training sets, or knowledge gained from radiologists. The weakness of rule-based systems is that the rules often lack necessary generality and are limited when multiple features are not independent. Partial volume effects and image noise may reduce the accuracy of hard thresholding. However, other classification schemes, such as a fuzzy segmentation method, would eventually require certain implicit or explicit threshold values for objective or cost functions to make decisions for the classification.

There are multiple limitations to our study. First, the number of cases is small. Although our preliminary CAD results are encouraging, we need more cases to test for generality. We will continue to collect CT cases of nodules of various sizes and shapes for additional training and testing. Alternatively, in the near future, we may be able to access the image database that is being developed by the Lung Image Database Consortium for the development, training, and evaluation of CAD methods (30). No extensive performance comparison between CAD and radiologists was performed in our study for evaluation of the role of CAD in clinical practice.

Second, the reference standard of nodules was defined by means of consensus reading. No histopathologic proof of nodules was obtained. Although interreader variability among radiologists in the definition and detection of nodules is likely substantial, it was not evaluated in this study. Since the nodules detected with CAD were included in the reference standard, the study results could be potentially biased toward making the CAD appear more accurate than it really is. At this point, we were unable to confirm or reject the influence of this potential bias on our study results.

Third, the size criteria of nodules were defined arbitrarily. The upper size limit of a nodule was based on the Fleischner Society's definition of a nodule—that is, a round opacity, at least moderately well marginated and no larger than 30 mm (21). The lower size limit of 3 mm was determined subjectively. Fourth, only CT cases of well-defined nodules were considered in our study. No ground glass opacity or alveolar infiltrative nodules were considered. The detection of these nonsolid nodules would require a modification of current CAD implementation and an introduction of new algorithms and parameters. The CT cases in our study did not contain areas of complex parenchymal disease. When other substantial abnormalities are present in the lung, our nodule detection scheme would need to be augmented with other detection methods to identify these abnormalities. Fifth, some of the cases in our study involved multiple nodules on the images. The multiplicity of nodules per case may introduce a bias in assessing the CAD performance. One approach to account for this problem is to compute and report the mean and standard deviation of the sensitivity and false-positive findings per case instead of the current approach of computing and reporting the sensitivity and false-positive findings for the pooled nodules divided by the number of cases.

In our study, the results of both approaches were similar. We do not believe that a large number of nodules per case would necessarily make the detection task easier for the CAD system than a small number of nodules per case. On the contrary, CAD would be useful in finding additional nodules in patients with a large number of nodules. In this respect, one practical approach to implementing the CAD system for the detection of pulmonary nodules would be to let the CAD system evaluate the thinly reconstructed images and perform a thorough search for nodules after a radiologist has already reviewed thicker sections (fewer images) to look at overall lung disease.

Sixth, the robustness of the rules in our nodule classification system was not evaluated. With the current implementation, no systematic study was performed to assess the relative weights and valid ranges of parameter values on the detection performance of the program. We plan to perform a rigorous feature selection analysis for a larger data set to evaluate the parameter and threshold values of the features that would provide maximum separation between the nodule and nonnodule objects. Furthermore, receiver operating characteristic analysis would be desirable to determine optimal operating points and to evaluate the ability of the classifier in the task of differentiating nodule candidates that correspond to actual nodules from false-positive candidates (13).

In summary, we developed an automated CAD program that takes advantage of thin-section volumetric data of multi–detector row CT images for the detection of pulmonary nodules. Our program was organized to systematically detect nodules of three different types (isolated, juxtapleural, and juxtavascular) in terms of their locations or adjacent anatomic structures. The results of our preliminary study demonstrated that a CAD system could detect pulmonary nodules, including small ones, with high sensitivity and a relatively low false-positive detection rate. Such a system may assist radiologists in the interpretation of CT images, particularly for lung cancer screening.

	FOOTNOTES

 

Abbreviations: CAD = computer-aided detection • 3D = three dimensional • 2D = two dimensional

² Current address: Korea Advanced Institute of Science and Technology, Daejeon, South Korea

³ Current address: Department of Radiology, Choongnam National University, Daejeon, South Korea

See also the article by Kim et al in this issue

Authors stated no financial relationship to disclose.

Author contributions: Guarantors of integrity of entire study, K.T.B., J.S.K., J.H.K.; study concepts and design, all authors; literature research, K.T.B., J.S.K., J.H.K.; clinical studies, K.T.B., J.H.K.; data acquisition and analysis/interpretation, all authors; statistical analysis, J.S.K., J.H.K.; manuscript preparation, definition of intellectual content, editing, revision/review, and final version approval, all authors

	References

TOP ABSTRACT INTRODUCTION Materials and Methods Results Discussion References

 

1. Heitzman ER. The role of computed tomography in the diagnosis and management of lung cancer: an overview. Chest 1986; 89(suppl 4):237S–241S.[Abstract/Free Full Text]
2. Howe MA, Gross BH. CT evaluation of the equivocal pulmonary nodule. Comput Radiol 1987; 11:61–67.[CrossRef][Medline]
3. Remy-Jardin M, Remy J, Giraud F, Marquette CH. Pulmonary nodules: detection with thick-section spiral CT versus conventional CT. Radiology 1993; 187:513–520.[Abstract/Free Full Text]
4. Fischbach F, Knollmann F, Griesshaber V, Freund T, Akkol E, Felix R. Detection of pulmonary nodules by multislice computed tomography: improved detection rate with reduced slice thickness. Eur Radiol 2003; 13:2378–2383.[CrossRef][Medline]
5. Henschke CI, McCauley DI, Yankelevitz DF, et al. Early Lung Cancer Action Project: overall design and findings from baseline screening. Lancet 1999; 354:99–105.[CrossRef][Medline]
6. Yankelevitz DF, Reeves AP, Kostis WJ, Zhao B, Henschke CI. Small pulmonary nodules: volumetrically determined growth rates based on CT evaluation. Radiology 2000; 217:251–256.[Abstract/Free Full Text]
7. Ko JP, Betke M. Chest CT: automated nodule detection and assessment of change over time—preliminary experience. Radiology 2001; 218:267–273.[Abstract/Free Full Text]
8. MacMahon H, Engelmann R, Behlen FM, et al. Computer-aided diagnosis of pulmonary nodules: results of a large-scale observer test. Radiology 1999; 213:723–726.[Abstract/Free Full Text]
9. Bae KT, Giger M, MacMahon H, Doi K. Computer-aided detection of pulmonary nodules in CT images (abstr). Radiology 1991; 181(P):144.
10. Giger M, Bae KT, MacMahon H. Computerized detection of pulmonary nodules in computed tomography images. Invest Radiol 1994; 29:459–465.[CrossRef][Medline]
11. Kanazawa K, Kawata Y, Niki N, et al. Computer-aided diagnosis for pulmonary nodules based on helical CT images. Comput Med Imaging Graph 1998; 22:157–167.[CrossRef][Medline]
12. Reeves AP, Kostis WJ. Computer-aided diagnosis for lung cancer. Radiol Clin North Am 2000; 38:497–509.[CrossRef][Medline]
13. Armato SG, III, Giger ML, MacMahon H. Automated detection of lung nodules in CT scans: preliminary results. Med Phys 2001; 28:1552–1561.[CrossRef][Medline]
14. Lee Y, Hara T, Fujita H, Itoh S, Ishigaki T. Automated detection of pulmonary nodules in helical CT images based on an improved template-matching technique. IEEE Trans Med Imaging 2001; 20:595–604.[CrossRef][Medline]
15. Wormanns D, Fiebich M, Saidi M, Diederich S, Heindel W. Automatic detection of pulmonary nodules at spiral CT: clinical application of a computer-aided diagnosis system. Eur Radiol 2002; 12:1052–1057.[CrossRef][Medline]
16. Qian J, Fan L, Wei G, et al. Knowledge-based automatic detection of multi-type lung nodules from multi-detector CT studies. SPIE 2002; 4684:689–697.[CrossRef]
17. Brown MS, Goldin JG, Suh RD, McNitt-Gray MF, Sayre JW, Aberle DR. Lung micronodules: automated method for detection at thin-section CT—initial experience. Radiology 2003; 226:256–262.[Abstract/Free Full Text]
18. Li Q, Katsuragawa S, Doi K. Computer-aided diagnostic scheme for lung nodule detection in digital chest radiographs by use of a multiple-template matching technique. Med Phys 2001; 28:2070–2076.[CrossRef][Medline]
19. Sonka M, Fitzpatrick JM, eds. Handbook of medical imaging: medical image processing and analysis. Bellingham, Wash: SPIE, 2000.
20. van den Boomgard R, van Balen R. Methods for fast morphological image transforms using bitmapped images. Computer Vision, Graphics, and Image Processing: Graphical Models and Image Processing 1992; 54:252–254.
21. Austin JH, Muller NL, Friedman PJ, et al. Glossary of terms for CT of the lungs: recommendations of the Nomenclature Committee of the Fleischner Society. Radiology 1996; 200:327–331.[Free Full Text]
22. Russ JC. The image processing handbook. Boca Raton, Fla: CRC, 2002.
23. Oppenheim AV, Willsky AS. Signals and systems. Englewood Cliffs, NJ: Prentice-Hall, 1997.
24. Bird RE. Professional quality assurance for mammography screening programs (letter). Radiology 1990; 177:587.[Free Full Text]
25. Metz CE, Shen JH. Gains in accuracy from replicated readings of diagnostic images: prediction and assessment in terms of ROC analysis. Med Decis Making 1992; 12:60–75.
26. Zheng B, Ganott MA, Britton CA, et al. Soft-copy mammographic readings with different computer-assisted detection cuing environments: preliminary findings. Radiology 2001; 221:633–640.[Abstract/Free Full Text]
27. Gurcan MN, Sahiner B, Petrick N, et al. Lung nodule detection on thoracic computed tomography images: preliminary evaluation of a computer-aided diagnosis system. Med Phys 2002; 29:2552–2558.[CrossRef][Medline]
28. Reeves AP, Kostis WJ. Computer-aided diagnosis of small pulmonary nodules. Semin Ultrasound CT MR 2000; 21:116–128.[CrossRef][Medline]
29. Brown MS, McNitt-Gray MF. Medical image interpretation. In: Sonka M, Fitzpatrick JM, eds. Handbook of medical imaging. Bellingham, Wash: SPIE, 2000; 399–445.
30. Armato SG 3rd, McLennan G, McNitt-Gray MF, et al. Lung image database consortium: developing a resource for the medical imaging research community. Radiology 2004; 232:739–748.[Abstract/Free Full Text]

This article has been cited by other articles:

				K. W. Lee, M. Kim, D. S. Gierada, and K. T. Bae Performance of a Computer-Aided Program for Automated Matching of Metastatic Pulmonary Nodules Detected on Follow-Up Chest CT Am. J. Roentgenol., November 1, 2007; 189(5): 1077 - 1081. [Abstract] [Full Text] [PDF]

				M. E. Baker, L. Bogoni, N. A. Obuchowski, C. Dass, R. M. Kendzierski, E. M. Remer, D. M. Einstein, P. Cathier, A. Jerebko, S. Lakare, et al. Computer-aided Detection of Colorectal Polyps: Can It Improve Sensitivity of Less-Experienced Readers? Preliminary Findings Radiology, October 1, 2007; 245(1): 140 - 149. [Abstract] [Full Text] [PDF]

				P. Peloschek, J. Sailer, M. Weber, C. J. Herold, M. Prokop, and C. Schaefer-Prokop Pulmonary Nodules: Sensitivity of Maximum Intensity Projection versus That of Volume Rendering of 3D Multidetector CT Data Radiology, May 1, 2007; 243(2): 561 - 569. [Abstract] [Full Text] [PDF]

				J.-S. Kim, J.-H. Kim, G. Cho, and K. T. Bae Automated Detection of Pulmonary Nodules on CT Images: Effect of Section Thickness and Reconstruction Interval--Initial Results Radiology, July 1, 2005; 236(1): 295 - 299. [Abstract] [Full Text] [PDF]

This Article Abstract Figures Only Full Text (PDF) All Versions of this Article: 2361041286v1 236/1/286    most recent Submit a response Alert me when this article is cited Alert me when eLetters are posted Alert me if a correction is posted Citation Map Services Email this article to a friend Similar articles in this journal Similar articles in PubMed Alert me to new issues of the journal Download to citation manager Citing Articles Citing Articles via HighWire Citing Articles via Google Scholar Google Scholar Articles by Bae, K. T. Articles by Kim, J.-H. Search for Related Content PubMed PubMed Citation Articles by Bae, K. T. Articles by Kim, J.-H.