| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
Technical Developments |
1 From the Department of Radiology, Massachusetts General Hospital, Boston (J.P.K.); and the Computer Science Department, Boston University, Boston, Mass (M.B.). Received January 20, 2000; revision requested March 3; revision received April 25; accepted May 1. Address correspondence to J.P.K., Department of Radiology, New York University Medical Center, 560 First Ave, New York, NY 10016 (e-mail: jane.ko@med.nyu.edu).
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMaterials and Methods ResultsDiscussionREFERENCES |
|---|
Index terms: Computed tomography (CT), computer programs • Computed tomography (CT), image processing • Computers, diagnostic aid • Lung, nodule, 60.281
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and Methods ResultsDiscussionREFERENCES |
|---|
To our knowledge, research about automated assessment of change in nodule number and size has not been performed previously. Studies have primarily addressed automated nodule detection at CT (2–5) and chest radiography (6,7). Findings demonstrate that computer-aided diagnosis systems help improve a radiologist’s receiver operating characteristic curve for detecting pulmonary nodules (8) and characterizing interstitial lung disease (9) on chest radiographs. However, there is a need to better quantify the size of pulmonary nodules and their change over time. To estimate change in nodule volume, the radiologist typically uses bidimensional measurements in the transverse plane. These measurements are subject to inter- and intraobserver variations, which can be decreased with a computer system. Yankelevitz et al (10) demonstrate that growth in malignant tumors as small as 5 mm can be detected at early repeat CT with use of a semiautomated method.
A computer system for automated nodule detection and quantification capabilities could be applied to a high-volume screening scenario such as low-dose CT, which has been studied as a screening test for lung cancer in a high-risk population (11). The failure to identify a nodule at low-dose CT (12) may crucially affect a patient’s prognosis, and a vision system could help decrease the number of missed nodules.
Given the potential applications to treatment and screening of patients with cancer, we developed a computer vision system that not only detects pulmonary nodules at CT but also quantifies their volume and change over time. The purpose of this study was to evaluate the performance of the computer vision system.
|
Materials and Methods |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and Methods ResultsDiscussionREFERENCES |
|---|
The chest helical CT studies (HiSpeed Advantage; GE Medical Systems, Milwaukee, Wis) were performed according to a standard departmental protocol: 1:1 pitch from the lung apices through the adrenal glands, with either 5-mm collimation (eight studies) or 10-mm collimation with 5-mm collimation through the hila (eight studies). Reconstruction intervals were 10 and 5 mm for imaging performed with 10- and 5-mm collimation, respectively. Of the 16 studies, eight were performed with intravenous administration of contrast material (Oxilan; Cook, Bloomington, Ind) and eight were performed without contrast material. Images were acquired with a 512 x 512 matrix and quantized with 16 bits. The CT data were transferred in the digital imaging and communications in medicine (DICOM) format to the computer on which the nodule detection computer system was implemented. A DICOM reader (ALICE, version 4.0; Hayden Image Processing Group, Parexel International, Boulder, Colo) was used to translate the CT data from DICOM format into binary data format.
Radiologist Analysis
One thoracic radiologist (J.P.K.), without knowledge of the results of the computer system, evaluated the CT scans for nodules. To maximize nodule detection, initial and follow-up scans were evaluated concurrently by using a clinical picture archiving and communication system workstation (Impax; Agfa, Ridgefield Park, NJ) with a 12-bit display. Nodules were identified and labeled on the hard copies of each CT study. Normal structures that were believed to simulate nodules were not marked. The radiologist noted nodule development, increase, stability, decrease, or resolution that occurred between the initial and follow-up studies. The radiologist estimated change visually unless there was uncertainty, at which point electronic calipers were used for clarification, particularly when studies were acquired with different fields of view. Nodules that were adjacent to a vessel or lung border were identified.
Computer Vision System
The computer vision system we developed was implemented on a personal computer (Dimension XPS, Dell Computer, Round Rock, Tex; LINUX version 5.2, Red Hat, Durham, NC; Pentium II, Intel, Phoenix, Ariz). A computer vision system is a collection of computer programs that performs automated analysis of image data (13).
Thorax and lung border detection.—Hounsfield units for attenuation were translated into brightness values during conversion into DICOM format. Thresholding of each CT image was performed to create a binary image. The higher attenuation soft tissue and bones were visible as bright gray values, whereas the lower attenuation air-filled lung parenchyma was visible as dark gray values. The thorax was detected by means of analysis of vertical and horizontal profiles of the binary image (Fig 1a). These profiles were used to identify an initial point on the lung border. Beginning at this point, the border was traced with a method based on a backtracking algorithm (14) that detected and bridged narrow channels that were vessels or artifacts. These channels were included in the lung parenchyma.
|
|
|
The centroid of an area was defined as the central balance point for the area. The centroids for the individual right and left lungs were calculated on the basis of the number and locations of pixels in the parenchyma of the right and left lungs, respectively (Fig 1c). The number and locations of pixels in both lungs were used to calculate the main centroid.
Determination of candidate regions.—The pixels in the lung parenchyma were divided into two subsets: a brighter subset, which related to normal structures (vessels, bronchi) and lung pathologic conditions (nodules), and a darker subset, which reflected the aerated lung. Structures that filled the entire section thickness appeared brighter than structures that were smaller than the section thickness, because the latter were averaged with the surrounding lung. To find nodules that were smaller than the section thickness, several gray-level thresholds were applied to create binary images that contained different candidate regions (Fig 2). The sequential labeling algorithm described by Horn (13) distinguishes and labels the candidate regions.
|
Shape determination was performed with two methods to characterize candidate regions as either nodules, which are circular, or vessels, which are elongated. However, circular regions on CT images may also be vessels in cross section. Lines of pixels for which the sum E of the square of the distance to points in the region were a minimum (Emin) and maximum (Emax) were automatically identified for each candidate region. The ratio Emin/Emax was 0 for a straight-line region and 1 for a circular region; therefore, the ratio is an indicator of how elongated or circular the region is. The second measure of shape made use of a circle surrounding the candidate region. A rounder candidate region occupied a larger percentage of pixels in the circle than did an elongated candidate region.
Analysis of consecutive CT sections with three-dimensional techniques.—The CT matrix location of a nodule centroid in one image was used as an estimate for the position of the projected nodule centroid in the subsequent image. A fixed area of 102 pixels around this projected centroid was searched automatically to locate the centroid of a corresponding nodule. On images that depicted the trachea, the centroids of the trachea were calculated and registered to align consecutive images in the same study. Small differences in the positions of tracheal centroids were used to adjust the search area for a corresponding nodule (Fig 3).
|
(d/2)3 and expressed in cubic millimeters. Analysis of change over time.—Identification of corresponding nodules on separate studies was more challenging than was identification of corresponding nodules on consecutive images in the same study. Differences in patient position and inspiration complicated registration between two different CT studies.
For each image in a patient’s initial CT study, the computer identified a possible matching image and two surrounding images in the follow-up CT study. This was performed with use of centroids of anatomic structures such as the sternum, vertebra, and trachea (Fig 3). By aligning anatomic centroids, differences in translation and rotation could be addressed. The centroid of the trachea was used because the trachea is typically a midline structure. However, the trachea can be shifted secondary to atelectasis or an adjacent mass and may not be a consistent landmark for registration. The most medial, lateral, anterior, and posterior pixels of each lung were also identified and registered as were the centroids of the individual and combined lungs. Results of global registration were used to quantify the translational and rotational differences between images. Given the centroid of a nodule in one image, a projected centroid of the same nodule in an image from a subsequent study was calculated with the translational and rotational parameters generated from the global registration of thoracic structures. A fixed area of 102 pixels around this projected centroid was searched to locate the centroid of the corresponding nodule. The initial study was similarly searched to identify nodules that corresponded with nodules identified on the follow-up study. The sizes of corresponding nodules in the two studies were then compared.
With the computer vision system used in this study, the lung apices needed to be identified manually on all studies. The computer system then searched for corresponding images between the two studies. In all patients, human intervention was needed to match studies obtained with differing section thickness and large variations in patient inspiration and to resolve ambiguities when the search in the local neighborhood resulted in identification of more than one possible corresponding nodule.
Computer Analysis of Data
The computer vision system analyzed each study and applied rule-based classification tests to identify "highly likely nodules," "likely nodules," and "normal structures." Results by the computer were compared with those by the radiologist. A Student t test was performed (EXCEL; Microsoft, Redmond, Wash) to identify any differences in nodule detection rates between CT studies performed with and those performed without intravenous contrast material. The computer system measured the diameters of both detected and missed nodules and placed nodules into three size categories. The categories, which were based on findings with pulmonary nodules at low-dose CT (4), were large (>10 mm), medium (
5 mm, 
10 mm), or small (<5 mm). The number of nodules in each category was counted, and the mean nodule size for each category and the entire study was tabulated.
The imaging computer system compared two sequential studies from the same patient and estimated changes in the size of a nodule by assessing for a change in size category. The relationship of assessments of nodule change by the computer and by the radiologist was determined with the Spearman rank correlation coefficient (15) (software, SAS, SAS Institute, Cary, NC; computer, Hewlett Packard, Palo Alto, Calif).
|
Results |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and Methods ResultsDiscussionREFERENCES |
|---|
|
On average, 76 candidate regions were classified for each image, and 3% (2.3 of 76) of these regions were misclassified as nodules on the basis of the threshold used. On average, 4.8% (12 of 252) of the candidate regions were misclassified with lower thresholds, and 3% (one of 33) were misclassified with higher thresholds. Nodule detection rates were 86% (199 of 230 nodules) for the eight nonenhanced studies and 85% (121 of 142) for the eight contrast-enhanced studies (difference not significant [P = .91]). Table 2 summarizes the size analysis with the computer. Sixty (16.2%) of the nodules were large; 123 (33.2%), medium; and 187 (50.5%), small (range, 0.7 mm to 3.6 cm).
|
|
|
|
Discussion |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and Methods ResultsDiscussionREFERENCES |
|---|
Although our computer system is still preliminary, our results for nodule detection are promising. The computer system identified 86% of all nodules and 91% of nodules larger than 3 mm. The missed nodules were predominantly smaller than 3 mm or contacted the lung border. To ensure that nodules abutting the lung border were not unintentionally excluded from the lung parenchyma, the computer system automatically analyzed the curvature of points along the lung border and inserted border segments into high-curvature portions of the border. The computer systems of Kanazawa et al (3) and Armato et al (2) also insert border segments when high-curvature points are detected. The system of Armato et al detects high-curvature points when a disk rolled along the initially detected border contacts the border at two points rather than one.
A computer system with the ability to register thoracic CT images between different studies has not been previously reported, to our knowledge. We have developed registration techniques and combined them with automated nodule recognition techniques to identify nodules on two studies and compare them.
Registration of studies is challenging owing to differences in rotation and translation of the imaged structures. Additional difficulties arise in the registration of thoracic images as a result of differences in patient inspiration. The radiology literature reports many registration techniques for the brain and other organ systems. For example, positron emission tomography (PET) has been correlated with magnetic resonance (MR) imaging for the brain (16–19). Nuclear medicine bone scans have been registered with bone radiographs (20). Preliminary studies have been performed for registration of PET and CT scans in the thorax (21). Registration methods often require some manual input to compensate for rotational and translational differences between two studies (16,18). For example, registration of PET and MR studies of the brain was aided by the manual identification of the midsagittal plane and subsequent reconstruction of the transverse images (17).
Our computer system accounted for differences in patient rotation and translation by computing centroids of the lungs, trachea, vertebra, and sternum. Our techniques, although not entirely automated, enable comparison of nodules in two different CT studies. Methods that automatically account for patient tilt in the craniocaudal dimension, differences in respiration, and alignment of the lung apices are challenging and need to be developed further. Our computer vision system can identify a nodule on multiple consecutive sections without human intervention if a nodule appears on more than one image in the same study. Further work will deal with the differentiation of a nodule from vertically oriented vessels.
The computer vision system automatically measured the diameter of the nodule and determined any change in nodule diameter between two studies. These measurements can be converted into volumetric values. Assuming a nodule is a sphere, a change in nodule diameter by 26% creates a doubling of nodule volume (22). In our study, for example, a spherical nodule had a diameter of six pixels, which translated into a diameter of 4.5 mm and a volume of 47.7 mm3. The computer vision system then detected the same nodule on the follow-up study and measured the diameter as seven pixels, or 5.25 mm. The volume was calculated as 75.7 mm3, which was a 62% increase. This example illustrates that a one-pixel difference in diameter translates into a large change in volume.
A one-pixel difference is difficult for a radiologist to ascertain; therefore, it is desirable for a computer system to automatically identify such a small pixel change. The one-pixel difference may not be due to a change in volume but may be secondary to sampling inaccuracies. Measurement of the apparent nodule size by the computer system is affected by the threshold values for pixel intensities that are used to distinguish brighter soft-tissue structures from the darker air-containing components of the lung. The CT scanner, kilovolt potential, reconstruction algorithm, section thickness, and nodule location in the field of view have been shown to affect pixel intensity (23) and, therefore, apparent nodule diameter. Further studies are needed to establish statistical bounds on the estimation error of change-in-diameter measurements. To avoid potential sampling errors in this preliminary study, we reported the change in nodule size determined by the computer in terms of change in size category. Although assessments of nodule change by the computer and radiologist correlated well with our choice of categories, a small decrease in nodule size may not have satisfied the criteria for change in category for the computer and therefore may account for some differences.
As in other preliminary studies (3,5), true- and false-positive rates reported in this study were computed for data that were used to develop the classification rules of the computer vision system. In the future, we will create a substantially larger database and separate it into training and test cases. Nodules in the training cases will be analyzed to improve the classification rules, and computer system efficacy will be assessed with test cases only and with more than one radiologist.
The computer vision system used in this study has quantitative abilities that could affect many facets of oncology care, from diagnosis of metastases to assessment of their response to chemotherapy. In the future, low-dose chest CT could be used for lung cancer screening. Use of this vision system could help ensure that nodules do not fail to be detected. In screening mammography, computer systems have been shown to improve receiver operating characteristic curves of radiologists for rating the likelihood of malignancy (24) and detecting abnormalities (25). Computer-aided diagnostic imaging systems that interact with a radiologist by means of an interface are being developed in mammography (26). Such an interface could easily be added to this computer vision system and would have many applications in thoracic radiology.
Our preliminary computer vision system demonstrates the potential for a clinically useful automated nodule detection system that quantifies nodule size and assesses for change over time. With further development, such a computer system could be applied to clinical scenarios in which objective nodule assessment is necessary.
|
ACKNOWLEDGMENTS |
|---|
|
FOOTNOTES |
|---|
Author contributions: Guarantors of integrity of entire study, J.P.K., M.B.; study concepts and design, J.P.K., M.B.; definition of intellectual content, J.P.K., M.B.; literature research, J.P.K., M.B.; experimental studies, J.P.K., M.B.; data acquisition, J.P.K., M.B.; data analysis, J.P.K., M.B.; statistical analysis, J.P.K., M.B.; manuscript preparation, editing, and review, J.P.K., M.B.
|
REFERENCES |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and Methods ResultsDiscussionREFERENCES |
|---|
This article has been cited by other articles:
|
|
 |
|
  D. Hock, R. Ouhadi, R. Materne, A.-S. Aouchria, I. Mancini, T. Broussaud, P. Magotteaux, and A. Nchimi Virtual Dissection CT Colonography: Evaluation of Learning Curves and Reading Times with and without Computer-aided Detection Radiology, September 1, 2008; 248(3): 860 - 868. [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]
|
|
|
|
|
|
C. Beigelman-Aubry, P. Raffy, W. Yang, R. A. Castellino, and P. A. Grenier Computer-Aided Detection of Solid Lung Nodules on Follow-Up MDCT Screening: Evaluation of Detection, Tracking, and Reading Time Am. J. Roentgenol., October 1, 2007; 189(4): 948 - 955. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 H Takao, I Doi, and M Tateno Evaluation of an automated system for temporal subtraction of thin-section thoracic CT Br. J. Radiol., February 1, 2007; 80(950): 85 - 89. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
H. P. McAdams, E. Samei, J. Dobbins III, G. D. Tourassi, and C. E. Ravin Recent Advances in Chest Radiography Radiology, December 1, 2006; 241(3): 663 - 683. [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]
|
|
|
|
|
|
L. R. Goodman, M. Gulsun, L. Washington, P. G. Nagy, and K. L. Piacsek Inherent Variability of CT Lung Nodule Measurements In Vivo Using Semiautomated Volumetric Measurements. Am. J. Roentgenol., April 1, 2006; 186(4): 989 - 994. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
K. Awai, K. Murao, A. Ozawa, Y. Nakayama, T. Nakaura, D. Liu, K. Kawanaka, Y. Funama, S. Morishita, and Y. Yamashita Pulmonary Nodules: Estimation of Malignancy at Thin-Section Helical CT--Effect of Computer-aided Diagnosis on Performance of Radiologists Radiology, April 1, 2006; 239(1): 276 - 284. [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]
|
|
|
|
|
|
C. L. Partain, H.-P. Chan, J. G. Gelovani, M. L. Giger, J. A. Izatt, F. A. Jolesz, K. Kandarpa, K. C. P. Li, M. McNitt-Gray, S. Napel, et al. Biomedical Imaging Research Opportunities Workshop II: Report and Recommendations Radiology, August 1, 2005; 236(2): 389 - 403. [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]
|
|
|
|
|
|
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]
|
|
|
|
|
|
S. G. Armato III, G. McLennan, M. F. McNitt-Gray, C. R. Meyer, D. Yankelevitz, D. R. Aberle, C. I. Henschke, E. A. Hoffman, E. A. Kazerooni, H. MacMahon, et al. Lung Image Database Consortium: Developing a Resource for the Medical Imaging Research Community Radiology, September 1, 2004; 232(3): 739 - 748. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
K. Awai, K. Murao, A. Ozawa, M. Komi, H. Hayakawa, S. Hori, and Y. Nishimura Pulmonary Nodules at Chest CT: Effect of Computer-aided Diagnosis on Radiologists' Detection Performance Radiology, February 1, 2004; 230(2): 347 - 352. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
S. G. Armato III, F. Li, M. L. Giger, H. MacMahon, S. Sone, and K. Doi Lung Cancer: Performance of Automated Lung Nodule Detection Applied to Cancers Missed in a CT Screening Program Radiology, December 1, 2002; 225(3): 685 - 692. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
| RADIOLOGY | RADIOGRAPHICS |
) to begin tracing of the lung border. (b) Diagram demonstrates border correction. A nodule that touches the lung border is initially excluded from the lung parenchyma. Border correction is performed by comparing the curvatures at points on the lung border, and the lung nodule is then included in the lung parenchyma. (c) On a chest CT image, the lung border is traced. The pixels of both lungs are used to calculate the main centroid (white cross marked M), whereas the pixels of the individual lungs are used to calculate the respective right and left lung centroids (white crosses labeled R and L). The black crosses indicate the most anterior, posterior, lateral, and medial pixels of each lung. Candidate regions detected by the computer that represent either nodule or normal structures are identified.
