| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Technical Developments |
1 From the Department of Radiology, David Geffen School of Medicine at UCLA, 10833 Le Conte Ave, Los Angeles, CA 90095-1721. Received October 18, 2001; revision requested January 11, 2002; revision received March 11; accepted May 7, 2002. Address correspondence to M.S.B. (e-mail: mbrown@mednet.ucla.edu).
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMaterials and Methods ResultsDiscussionAPPENDIXREFERENCES |
|---|
3 mm). With the same images, a radiologist detected nodules and micronodules with sensitivities of 91% and 51%, respectively, without system input. With assistance from the automated system, these sensitivities increased to 95% and 74%, respectively. Preliminary results indicate that the automated system considerably improved the radiologist’s performance in micronodule detection. © RSNA, 2002
Index terms: Computed tomography (CT), computer programs • Computed tomography (CT), image processing • Computers, diagnostic aid • Lung, nodule, 60.281
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and Methods ResultsDiscussionAPPENDIXREFERENCES |
|---|
The goal of lung cancer screening with CT is the detection of small cancers, presumably when they are biologically early in their evolution and amenable to surgical cure. CT has become the focus of lung cancer screening strategies because it combines excellent contrast resolution with tomographic perspective, both of which enhance the detection of lung nodules while maintaining reasonably low radiation exposure. CT has been shown to be more sensitive than projectional radiography in the detection of lung nodules, particularly small nodules (5,6,7). The focus of the present study was detection of micronodules (<3 mm in diameter) to facilitate early diagnosis of lung cancer.
The original low-dose CT screening protocols for lung cancer were based on single-section helical CT technology, which generated 10-mm-thick images. The commercial availability of multi–detector row CT has made the acquisition of data sets with high spatial resolution practical, which allows consistent detection of micronodules. With multi–detector row CT, the chest can be surveyed in one suspended breath hold with high in-plane resolution, narrow collimation, and narrow reconstruction intervals. Nearly isotropic high-resolution data sets consistently render the fine complex bronchovascular anatomy of the peripheral lung in three dimensions, which captures the continuity of the arborizing of structures within and through the transverse imaging plane. This degree of anatomic precision is necessary if computerized systems are to discriminate accurately between normally branching anatomy and micronodular lung cancers. However, anatomic detail is achieved at the expense of large data volumes (7,8). Interpretation times for large thin-section data sets would be impractical if CT screening for lung cancer becomes routine; therefore, development of computer-assisted nodule detection schemes is a prerequisite.
We developed a computer system for automatic detection of lung micronodules in helical thin-section CT chest data sets. Basic CT segmentation methods, which use primarily attenuation thresholding, are insufficient for nodule detection because nodules and vascular structures have similar attenuation at CT. The system described herein augments conventional image processing tools by introducing anatomic knowledge relating to the expected size, shape, and location of nodules and lung anatomy to guide the segmentation process. Previously developed automated lung nodule detection systems were designed for use with thick-section CT data and were tested primarily with nodules larger than 5 mm (9,10). The system described herein represents a shift from thick- to thin-section CT and from nodule to micronodule detection.
The purpose of this study was to evaluate nodule and micronodule detection with a radiologist alone, the automated detection system alone, and the radiologist with the computer-aided system.
|
Materials and Methods |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and Methods ResultsDiscussionAPPENDIXREFERENCES |
|---|
According to clinical practice, individual volumetric images were acquired (HiSpeed CT/i; GE Medical Systems, Milwaukee, Wis) with technical parameters of 120 kV, 60 mA, 1-mm beam collimation, 2-mm table incrementation, 0.8 seconds per revolution, and reconstruction interval of 0.5–1.0 mm. Images were limited typically to a 20-mm longitudinal extent. Thin-section CT images were used since they approximate the technical factors that can be applied at multi–detector row CT in one breath hold.
Nodule Detection System
The nodule detection system automatically identifies and highlights lung nodules in volumetric CT data sets for confirmation or rejection by a user. The system recognizes anatomic landmarks and lung nodules with a two-step process. Image segmentation is performed to subdivide the volumetric image data into three-dimensional regions of interest. Then, object recognition is performed in which the regions are labeled anatomically. Simple attenuation thresholding alone is insufficient for this task because of the similar attenuation properties of nodules, blood vessels, and soft-tissue anatomy. Region growing on the basis of only attenuation does not allow discrimination between them; therefore, the nodule detection system makes use of additional knowledge of anatomy to recognize objects in an image.
The nodule detection system uses three-dimensional segmentation involving attenuation thresholding, region growing (11), and mathematical morphology (12) to identify regions of interest in a given CT data set. The system then labels these regions on the basis of a model of lung nodules and relevant intrathoracic anatomy (13,14). The model describes the expected size, shape, topology, and x-ray attenuation of these structures. In addition to lung nodules, the model describes blood vessels and distinguishes between the two, which is critical in nodule identification. To accomplish this task, three-dimensional shape information is important since nodules tend to be approximately spherical, whereas bronchovascular structures are more tubular. The description in the model is flexible enough, however, to provide automatic discrimination of both regular and irregular (spiculated) nodules from bronchovascular anatomy. The system also uses regionally varying size information since vessels decrease in size toward the lung periphery, while nodules may be larger in cross section and volume.
From the 29 data sets collected, a subset of 14 was selected at random for use in model development. These cases were used to develop the descriptions and set parameters contained in the model. The remaining 15 image data sets were kept hidden from the system until the evaluation phase.
When the system is applied to a CT data set, it performs automated nodule detection, and regions labeled as nodules are highlighted on the images for review by a user (Figs. 1–3). Further details of the modeling, segmentation, and object recognition techniques are provided in the Appendix.
|
|
|
|
|
|
|
3 mm) and nodules (>3 mm). The nodule detection system was applied to the 15 evaluation CT data sets (Ultra 10 workstation [with UNIX operating system]; Sun Microsystems, Palo Alto, Calif). To assess the performance of the automated nodule detection system by itself, the data sets were input to the system. The system segmented anatomic structures and identified candidate nodules. These candidate nodules were then compared on the basis of location, with the nodules identified by the truth committee.
A second experiment was performed to assess the usefulness of the system as a diagnostic adjunct to the visual review of data sets. The same 15 evaluation data sets were reviewed in blinded fashion by a thoracic radiologist (R.D.S.) with 6 years of experience who was asked to identify the number and location of nodules. This radiologist was not part of the truth committee and did not perform the original interpretation of the images. The images were reviewed with two sequential conditions. The first condition involved visual review of the images in a soft-copy environment. The second condition was that the original images were viewable and candidate nodules were highlighted by the system (both true- and false-positive nodules). Differences in performance were measured between the two reading conditions. In this initial trial, there was no effort made to determine differences in the time required for interpretation with and that without use of the automated system.
Statistical Evaluation
The performance of the system was evaluated in terms of sensitivity for nodule detection and the number of false-positive findings per case. These measures were also used to compare the performance of the radiologist with use of the two reading conditions. Results were stratified on the basis of nodule size into micronodules (<3 mm) and nodules (>3 mm).
|
Results |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and Methods ResultsDiscussionAPPENDIXREFERENCES |
|---|
The automated system, without any user interaction, achieved a sensitivity of 100% (22 of 22) for nodules and 70% (40 of 57) for micronodules, with an average of 15 false-positive findings per subject data set (total, 219 false-positive findings). These false-positive findings came from the following main sources: (a) 52 from vessels (both blood vessels and airway walls) that were truncated in the longitudinal (z-axis) direction owing to the limited thin-section sequences; (b) 45 from blood vessels (not truncated); (c) 27 from airway walls (not truncated); (d) 18 from areas of parenchymal disease that resulted in ground-glass, reticular, or airspace attenuation; and (e) 77 from other sources.
The radiologist independently (without system input) detected nodules with a sensitivity of 91% (20 of 22) and micronodules with a sensitivity of 51% (29 of 57), with 0.3 false-positive finding per data set.
With the additional input provided by the automated system, sensitivity for the radiologist increased to 95% (21 of 22) of nodules and 74% (42 of 57) of micronodules. The number of false-positive nodules remained unchanged. These results are summarized in the Table. Figure 4, A and B illustrate two examples in which the automated system detected micronodules that were initially missed by the radiologist, who used visual inspection.
|
|
|
Discussion |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and Methods ResultsDiscussionAPPENDIXREFERENCES |
|---|
A small number of knowledge-based schemes are under development to automate lung nodule detection at CT, and results are promising (9,10,17,18). These systems have been developed for thick-section (5–10 mm) imaging and have been designed and tested with lesions that are typically 1 cm or more in diameter (<5 mm, volume averaging from the thick sections makes automated detection difficult). These systems have been tested with limited numbers of cases and have achieved sensitivities of approximately 90%, averaging roughly 20 false-positive findings per case. Although these systems are promising, the logical extension of their use to include the detection of micronodules smaller than 3 mm in diameter is challenging, particularly in terms of the increased number of false-positive findings caused by small bronchovascular anatomy. The system used in the present study is particularly well suited to the handling of complexities associated with the distinction of micronodules from fine bronchovascular structures since the model can readily be extended to describe detailed anatomy. The modular system architecture also allows additional segmentation algorithms to be incorporated as required.
The results of the present study are preliminary since the system is still under development. It was tested with a relatively small number of cases with use of data from single-section scanners with which only limited thin-section sequences were performed through regions of suspected disease. Also, "truth" was determined visually by a committee, rather than on the basis of pathologic findings. Moreover, the effect on reader performance was assessed for only one reader. However, these results suggest that computer programs may favorably affect micronodule detection at CT. The results are clinically important in terms of motivating continued research in this area and in lung micronodule detection in particular.
The results suggest that the radiologist was less sensitive to micronodules than were the truth committee members when they evaluated images independently. There is the potential for high interobserver variability because readers were asked to mark nodules without ascribing clinical importance, which is not their usual practice. Since the truth committee members performed consensus reading during the course of the study, they had the opportunity to learn from each other and become consistent in terms of which opacities they called nodules. The radiologist did not receive any guidance from the truth committee and also had less experience as a thoracic radiologist. The results suggest that it is difficult even for experienced radiologists to agree whether an opacity should be considered a nodule. Criteria for labeling opacities as nodules are evolving in ongoing studies of lung cancer screening and computer-aided diagnosis.
The system achieved reasonable detection accuracy despite a crude anatomic model, a success that we ascribe in part to the richness of three-dimensional characterization achieved with thin-section image data. These major sources of false-positive nodules were presented in Results.
Truncated vessels.—Vessels that were truncated in the z axis owing to the limited thin-section sequences had nodulelike shape properties. This source of error can be expected to decrease with data sets that span the entire thorax.
Fragmented blood vessels.—Blood vessels that were fragmented as a result of volume averaging were segmented into regions that appeared like small nodules. This source of error will require the development of more sophisticated segmentation algorithms that rely less on fixed attenuation thresholds, which fail when the attenuation decreases below the threshold as a result of volume averaging. For example, multiple thresholds (18) or thresholds that vary adaptively as a vessel is tracked could be used.
Fragmented airway walls.—Some airway walls were fragmented as a result of volume averaging. A common source of misclassification by the system was the walls of the segmental bronchi. The model contained descriptions of pulmonary vessels but did not at the time contain descriptions of the airways. Part of the strength of the system is the ability to visualize system errors relative to existing anatomic rules and to add additional rules that circumvent repeated errors of misclassification. We expect to reduce misclassification of bronchial wall anatomy in subsequent system revisions.
Parenchymal disease.—Areas of parenchymal disease resulted in ground-glass, reticular, or airspace attenuation. Methods are under development to identify automatically various types of parenchymal disease (19,20). Because we are using an anatomic region-based approach to nodule detection, we may be able to alter the criteria for detecting a nodule within a region of complicated parenchymal disease to lessen the number of false-positive findings.
Reduction of the number of false-positive findings will be a focus of ongoing developments since full-chest multi–detector row CT will result in many more image sections than those in the data sets used in the present study. We do not expect a proportional increase in the number of false-positive findings, however, because multi–detector row CT will provide a smaller effective section thickness and less partial volume effect.
As multi–detector row systems make thin-section CT more routine in the thorax, more lung micronodules will be visible, either prospectively or retrospectively. An automated detection system has the potential to increase the number of prospective detections of small lesions, and a size threshold for reporting lesions may be required. The lower boundary of lesion size to be reported will depend on what is found to be clinically important in lung cancer screening trials currently underway. In fact, systems that assist in the detection of early micronodular cancers may be necessary to fully answer questions about the utility of lung cancer screening (eg, to determine whether changes in size at the subcentimeter level translate into predictable differences in the biologic behavior of lung cancer).
The majority of subjects included in the present study had lung, renal cell, and colon cancer with metastatic disease. Although an important potential application of this system is in screening for primary lung cancer, the nodules included in this preliminary test have similar appearance in terms of the characteristics used by the system for detection. The increased number of nodules in subjects with metastatic disease does not make the detection task easier for the system.
The results of this preliminary study suggest that a model-based system can detect lung nodules with high sensitivity and may also be reasonably accurate for the detection of small micronodules. The system can provide interpretation assistance to a radiologist reviewing large CT data sets. We are continuing to refine our model of thoracic anatomy and lung nodules to reduce misclassification of thoracic anatomy for nodules. Data collection is underway for a larger database of cases from a multi–detector row CT scanner that will provide high-resolution data sets through the entire thorax and enable more comprehensive development and evaluation of the automated nodule detection system.
|
APPENDIX |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and Methods ResultsDiscussionAPPENDIXREFERENCES |
|---|
Anatomic Modeling
The model describes anatomic-pathologic structures, such as lungs, pulmonary vessels, and nodules. The model is represented by using a semantic network (22). Each node in the model consists of an anatomic name and a set of features (described later). The arcs in the network represent structural relationships (also features) between anatomic objects. A simplified overview of the model is shown in Figure 1.
Each node contains the expected values for the following key features of an object: (a) x-ray attenuation range in Hounsfield units; (b) relative location based on relational features in the semantic network (eg, "part of," "inside"); (c) volume (v), in cubic millimeters, calculated for a segmented object, represented by a set of voxels, S, by summing the volume of each voxel contained in S:
|
|
|
To allow for anatomic variations, each numeric feature parameter has an allowable range of values represented with a fuzzy set (23). The fuzzy set provides a membership function that assigns a score between 0.0 and 1.0 to each possible feature value. A score of 0.0 indicates nonmembership in the fuzzy set of possible values, a score of 1.0 represents full membership, and a value in between indicates a degree of partial membership. Figure 2 shows an example of a fuzzy set that describes the possible range of values for the sphericity of a nodule. The membership function has value 0.0 for sphericity values of less than 5, since low sphericity is considered likely to indicate a blood vessel rather than a nodule. Conversely, sphericity values higher than 25 are given a membership score of 1.0, while bearing in mind that nodules can be spiculated (ie, not particularly spherical). The fuzzy sets for each feature in the model were determined empirically by reviewing a development set of 14 cases (selected at random from the available set of 29 cases) with the supervision of thoracic radiologists. A small development set such as this does not cover the range of possible anatomic and pathologic variations. It merely provides assistance in the process of extracting and quantifying the knowledge of the participating radiologists.
Detection of Lung Regions
The model divides the aerated lung into transverse and peripheral regions. The peripheral region is a 20-mm-wide strip around the outside of the lungs. Since vessels are smaller in the lung periphery, the system can be more aggressive in detecting nodules without incurring many false-positive findings.
Mathematical morphology (12) is used to compute transverse and peripheral lung regions. The transverse portion of the lung is determined on the basis of morphologic erosion of the lungs, as shown by the part of (eroded) relationship in Figure 1. The Peripheral Lung is then part of the Lung and not part of the Transverse Lung. Segmentation of transverse and peripheral lung regions is shown in Figure 3.
Nodule Detection
Nodules are detected from among the opacities (nodules and bronchovascular structures), modeled as inside the lung. Morphologic closing is applied to the lung so that nodules that contact the chest wall are enclosed. Transverse and peripheral nodules are modeled as part of the opacities (Fig 3, D). Since these nodules may be in contact with blood vessels, morphologic opening is applied to the opacities to separate nodule candidates from such vessels. A small structuring element is used for peripheral nodules since vessels are small in the periphery; a larger structuring element is used in transverse regions to allow nodules to be separated from large bronchovascular structures near the hilum. The segmentation process generates multiple image primitives (regions) that become candidates for matching to transverse or peripheral nodules as defined in the model (Fig 1).
The model describes opacities inside the lungs that are typically blood vessels or nodules; distinction between these is critical in nodule identification. To accomplish this, three-dimensional shape information (sphericity feature) is important since nodules are approximately spherical and bronchovascular structures are more tubular. Fuzzy logic is used to match opacities to either nodules or vessels. A confidence score, Ca, is calculated for each candidate, a, as follows:
|
(a) is a function that calculates the value of the feature (eg, sphericity or volume) for a given image region (nodule candidate), and M is the fuzzy membership function for the feature defined in the model (Fig 2). This fuzzy membership function, M, generates a confidence score in the range of 0.0–1.0. Regions matched to the Nodule object in the model with overall confidence of C > 0.9 were presented to the user as nodule candidates for confirmation or rejection. |
FOOTNOTES |
|---|
|
REFERENCES |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and Methods ResultsDiscussionAPPENDIXREFERENCES |
|---|
This article has been cited by other articles:
|
|
 |
|
  F. Girvin and J. P. Ko Pulmonary Nodules: Detection, Assessment, and CAD Am. J. Roentgenol., October 1, 2008; 191(4): 1057 - 1069. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
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. Das, G. Muhlenbruch, A. H. Mahnken, T. G. Flohr, L. Gundel, S. Stanzel, T. Kraus, R. W. Gunther, and J. E. Wildberger Small Pulmonary Nodules: Effect of Two Computer-aided Detection Systems on Radiologist Performance. Radiology, November 1, 2006; 241(2): 564 - 571. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
R. Yuan, P. M. Vos, and P. L. Cooperberg Computer-Aided Detection in Screening CT for Pulmonary Nodules. Am. J. Roentgenol., May 1, 2006; 186(5): 1280 - 1287. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
F. Li, H. Arimura, K. Suzuki, J. Shiraishi, Q. Li, H. Abe, R. Engelmann, S. Sone, H. MacMahon, and K. Doi Computer-aided Detection of Peripheral Lung Cancers Missed at CT: ROC Analyses without and with Localization Radiology, November 1, 2005; 237(2): 684 - 690. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
K. G. Kim, J. M. Goo, J. H. Kim, H. J. Lee, B. G. Min, K. T. Bae, and J.-G. Im Computer-aided Diagnosis of Localized Ground-Glass Opacity in the Lung at CT: Initial Experience Radiology, November 1, 2005; 237(2): 657 - 661. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
J. K. Leader, T. E. Warfel, C. R. Fuhrman, S. K. Golla, J. L. Weissfeld, R. S. Avila, W. D. Turner, and B. Zheng Pulmonary Nodule Detection with Low-Dose CT of the Lung: Agreement Among Radiologists Am. J. Roentgenol., October 1, 2005; 185(4): 973 - 978. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
K. T. Bae, J.-S. Kim, Y.-H. Na, K. G. Kim, and J.-H. Kim Pulmonary Nodules: Automated Detection on CT Images with Morphologic Matching Algorithm--Preliminary Results Radiology, July 1, 2005; 236(1): 286 - 293. [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]
|
|
|
|
|
|
J. M. Goo, T. Tongdee, R. Tongdee, K. Yeo, C. F. Hildebolt, and K. T. Bae Volumetric Measurement of Synthetic Lung Nodules with Multi-Detector Row CT: Effect of Various Image Reconstruction Parameters and Segmentation Thresholds on Measurement Accuracy Radiology, June 1, 2005; 235(3): 850 - 856. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 J. L. Mulshine New Developments in Lung Cancer Screening J. Clin. Oncol., May 10, 2005; 23(14): 3198 - 3202. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 K Doi Current status and future potential of computer-aided diagnosis in medical imaging Br. J. Radiol., January 1, 2005; 78(suppl_1): S3 - s19. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
G. D. Rubin, J. K. Lyo, D. S. Paik, A. J. Sherbondy, L. C. Chow, A. N. Leung, R. Mindelzun, P. K. Schraedley-Desmond, S. E. Zinck, D. P. Naidich, et al. Pulmonary Nodules on Multi-Detector Row CT Scans: Performance Comparison of Radiologists and Computer-aided Detection Radiology, January 1, 2005; 234(1): 274 - 283. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 D R Baldwin, J D Birchall, R H Ganatra, and K S Pointon Evaluation of the solitary pulmonary nodule: clinical management, role of CT and nuclear medicine Imaging, October 1, 2004; 16(1): 22 - 36. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
D. Gur, B. Zheng, C. R. Fuhrman, and L. Hardesty On the Testing and Reporting of Computer-aided Detection Results for Lung Cancer Detection Radiology, July 1, 2004; 232(1): 5 - 6. [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
