Radiology
HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS
QUICK SEARCH:   [advanced]


     

Published online before print September 28, 2005, 10.1148/radiol.2372041461
This Article
Right arrow Abstract Freely available
Right arrow Figures Only
Right arrow Full Text (PDF)
Right arrow All Versions of this Article:
2372041461v1
237/2/657    most recent
Right arrow Submit a response
Right arrow Alert me when this article is cited
Right arrow Alert me when eLetters are posted
Right arrow Alert me if a correction is posted
Right arrow Citation Map
Services
Right arrow Email this article to a friend
Right arrow Similar articles in this journal
Right arrow Similar articles in PubMed
Right arrow Alert me to new issues of the journal
Right arrow Download to citation manager
Right arrow reprints & permissions
Citing Articles
Right arrow Citing Articles via HighWire
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Kim, K. G.
Right arrow Articles by Im, J.-G.
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by Kim, K. G.
Right arrow Articles by Im, J.-G.
(Radiology 2005;237:657-661.)
© RSNA, 2005

Technical Developments

Computer-aided Diagnosis of Localized Ground-Glass Opacity in the Lung at CT: Initial Experience1

Kwang Gi Kim, MS, Jin Mo Goo, MD, Jong Hyo Kim, PhD, Hyun Ju Lee, MD, Byung Goo Min, PhD, Kyongtae T. Bae, MD, PhD and Jung-Gi Im, MD

1 From the Department of Radiology, Seoul National University College of Medicine, and the Institute of Radiation Medicine, SNUMRC, 28 Yongon-dong, Chongno-gu, Seoul 110-744, Korea (K.G.K., J.M.G., J.H.K., H.J.L., J.G.I.); Department of Biomedical Engineering, Seoul National University College of Medicine, Seoul, Korea (B.G.M.); and Mallinckrodt Institute of Radiology, Washington University School of Medicine, St Louis, Mo (K.T.B.). Received August 23, 2004; revision requested October 28; revision received November 19; accepted December 24. Supported in part by Seoul National University Hospital Research Grant (0420030300) and by a grant from the Korea Health 21 Research and Development Project-Ministry of Health and Welfare, Republic of Korea (03-PJ1-PG10-51300-0006). Address correspondence to J.M.G. (e-mail: jmgoo{at}plaza.snu.ac.kr).


   ABSTRACT
TOP
 ABSTRACT
INTRODUCTION
 Materials and Methods
 Results
 Discussion
 Appendix
 References
 
The purpose of this study was to develop an automated scheme to facilitate detection of localized ground-glass opacity (GGO) in the lung at computed tomography (CT). Institutional review board approval and informed consent were not required. Two radiologists reviewed CT images from 14 patients (five men, nine women) who had lung cancer or metastasis and whose malignancy was classified as GGO. The lung region was sampled and completely covered with contiguous, 50% overlapping regions of interest (ROIs) measuring 30 x 30 pixels in size. The lung area within each ROI was analyzed to compute texture features and gaussian curve fitting features. Performance of the artificial neural networks (ANNs) measured by using the area under the receiver operating characteristic curve was 0.92. With a threshold of 0.9, the sensitivity of the ANN for detecting GGO ROIs was 94.3% (280 of 297 ROIs), and the positive predictive value was 29.1% (280 of 963 ROIs). A computerized scheme may hold promise in facilitating detection of localized GGO at CT.

© RSNA, 2005


   INTRODUCTION
TOP
 ABSTRACT
 INTRODUCTION
Materials and Methods
 Results
 Discussion
 Appendix
 References
 
The goal of lung cancer screening with computed tomography (CT) is to detect small cancers, presumably when they are early in their biologic evolution and amendable to surgical cure. Previously, computer-aided diagnosis (CAD) systems were reported to be effective in facilitating detection of small pulmonary nodules at CT (1–7).

At CT, lung cancer may appear as either a solid nodule, a nodule with ground-glass opacity (GGO) (part-solid nodule), or a nodule with pure localized GGO (nonsolid nodule). Although solid nodules are the most common manifestation of lung cancer, nodules with GGO represent approximately 20% of the total nodules demonstrated at screening and have a higher malignancy rate than solid nodules (8). Data from many studies suggest that localized GGO represents an early or precursor to adenocarcinoma (9–13). Despite their potential clinical importance, nodules with GGO or localized GGO may not be detected at screening CT. According to a study conducted by Li et al (14), 27 of 39 lung cancers that were missed by radiologists at screening CT were nodules with GGO.

CAD programs may help improve the detection of nodules with GGO or localized GGO. Most CAD schemes that are used to facilitate detection of focal lung lesions, however, are designed and optimized for solid nodules. In one study (2), lung cancers that were missed at screening were detected by using the computerized method, which had a sensitivity of 84% (32 of 38 nodules); all six lung cancers that were missed by using the automated nodule detection program were either pure GGO (n = 4) or mixed GGO (n = 2). Although the detection and characterization of diffuse GGO by using automated computerized schemes have been previously described (15–18), to our knowledge researchers have yet to design and develop a computerized scheme specifically to facilitate detection of localized GGO. We postulate that an efficient CAD program can be developed by combining the texture and pixel attenuation features of localized GGO with an artificial neural network (ANN) classification. Thus, the objective of this study was to develop an automated scheme specifically to facilitate detection of localized GGO in the lung at CT.


   Materials and Methods
TOP
 ABSTRACT
 INTRODUCTION
 Materials and Methods
Results
 Discussion
 Appendix
 References
 
Patients
CT images that demonstrated localized GGO in 14 patients (five men, nine women; age range, 49–80 years; mean age, 63 years) were analyzed. Twelve of these patients had primary lung lesion(s) and underwent biopsy. Lesions were subsequently diagnosed as primary lung neoplasms and included nine adenocarcinomas, two bronchioloalveolar carcinomas, and one atypical carcinoid. When multiple regions of localized GGO were present in a patient, biopsy was performed in only one region of the localized GGO for diagnosis. Two patients were considered to have nonpulmonary cancers, including one leiomyosarcoma and one hepatocellular carcinoma. Because their pulmonary lesions were presumed to be metastasis, these two patients did not undergo biopsy. These patients also had solid nodules, which were not included in analysis. Our institutional review board does not require approval or patient informed consent for retrospective review of previously obtained image data. Patient confidentiality, however, was protected.

CT Image Acquisition and Evaluation
CT images were acquired by using one of three CT scanners (Somatom Plus 4, Siemens, Erlangen, Germany; LightSpeed Ultra, GE Medical Systems, Milwaukee, Wis; or Mx8000, Philips Medical System, Andover, Mass). Because data were collected retrospectively, a variety of scanning protocols were used, including thin-section CT performed at 1.0-cm intervals (nonhelical high-resolution CT) (n = 2), volumetric helical CT (n = 12), and CT with (n = 3) or without (n = 11) intravenous contrast material. Tube voltage ranged from 120 to 140 kV, and tube current ranged from 200 to 400 mA. The image matrix size was 512 x 512 pixels. The section thickness was 1.0 mm in eight patients, 2.0 mm in three patients, and 5.0 mm in three patients. The field of view was adjusted to optimize for the size of the patients and ranged from 300 to 350 mm, which resulted in a pixel size of 0.586–0.684 mm.

Two chest radiologists (H.J.L. and J.M.G., with seven and 13 years experience reading CT images of the chest, respectively) identified locations of GGO on CT images by consensus. GGO was defined as an area of increased attenuation without obscuration of the underlying vessels and bronchi (19). Images were displayed by using a lung window with a center of –700 HU and a width of 1500 HU. A total of 1524 images (range, 17–325 images) from 14 patients were reviewed. When there were multiple CT examinations performed per patient, the results from only one CT examination performed prior to biopsy were reviewed. For computerized analysis, 73 images that demonstrated localized GGO were selected by two authors (H.J.L., J.M.G). When an opacity that was recognized as GGO by the radiologists spanned more than one CT section, all involved CT sections were included. Because seven patients had multiple (up to five) regions of localized GGO, the total number of localized GGO regions was 29. One radiologist (J.M.G.) outlined the boundary of each localized GGO on soft-copy CT images by using a graphic user interface that was developed in-house. This outlined boundary served as the reference standard. The lesions ranged from 5 to 28 mm in diameter (mean, 15 mm; median, 13 mm) and from –740 to –174 HU in mean attenuation (mean, –435 HU; median, –387 HU).

Analysis of Image Features
CT data were transferred in the Digital Imaging and Communications in Medicine format to a personal computer (1-GHz Pentium processor with 256 MB of random access memory) for postprocessing. The environment is developed and programmed by using a computer programming language (Visual C++; Microsoft, Redmond, Wash). For each CT image, the lung regions were segmented by using a threshold and labeling technique (5). The segmented lung region was sampled and completely covered with contiguous, 50% overlapping regions of interest (ROIs) that were 30 x 30 pixels in size. This ROI size was determined on the basis of data from similar studies on the regional analysis of the lung at CT (16,17). When an area outside of the lung region was included in the ROI, the corresponding outside area was ignored for GGO evaluation. In this way, 17 734 ROIs from 73 images were generated. When an area of GGO occupied more than 50% of the ROI, the ROI was categorized as a GGO ROI. When an area of GGO occupied less than 50% of the ROI, the ROI was categorized as a non-GGO ROI.

Of the many texture features that have been described in the literature (20–22), five were used in our study, including kurtosis, surface curvature, and three gray-level co-occurrence matrices features—that is, inertia, maximum probability, and momentum. Kurtosis is defined as the extent to which a histogram is peaked; histograms with peaks that are sharper than those of a normal distribution have a positive kurtosis. Inertia, maximum probability, and momentum are features that describe the spatial dependence of gray-scale distributions and are derived from the set of gray-level co-occurrence matrices computed at each ROI (20). Surface curvature is the rate of change of the slope of the tangent to the surface (21). In addition to these texture features, features that were associated with gaussian curve fitting were computed according to the attenuation histogram of pixels within each ROI (see Appendix for details). Because of the characteristic pixel attenuation distribution of GGO, the use of gaussian curve fitting features enabled the differentiation of GGO from background pulmonary parenchyma.

ANN and Analysis for GGO Detection
A three-layered ANN with a back-propagation algorithm was used in this study. For each ROI, values for texture features and area under the curve of gaussian curve fitting were calculated and used as the input data for the ANN. Five texture features and the curve fitting operation feature were used as input data for the ANNs.

All input data were normalized to a range of 0–1.0. The number of hidden units was determined by selecting the most efficient number, starting from five. A nonlinear, sigmoid function was used as a transfer function for each of the neurons in the hidden and output layers of the networks. There was a single output node for the classification of ROIs as either positive or negative for GGO. The output value of the neural network was either 0 or 1. To obtain this output value, various thresholds were applied, which were automatically generated by using a commercially available software program (NeuroSolution; NeuroDimension, Gainesville, Fla) to plot the receiver operating characteristic (ROC) curve. When the output value of the ROI was greater than the threshold value, the system classified the ROI as a GGO ROI (ie, output value of 1). Conversely, when the output value of the ROI was less than the threshold value, the system classified the ROI as a non-GGO ROI (ie, output value of 0). Therefore, the numbers of input, hidden, and output units for the ANN were six, 10, and one, respectively.

The performance of the ANN was evaluated by using an ROC analysis (23,24) and the k-fold cross-validation method (NeuroSolution). For this method, data that are used for the training of the ANN are also used for testing. Binormal ROC curves were estimated with the ROCKIT algorithm (C. E. Metz, University of Chicago, Ill) by entering the ANN outputs, which were continuous values, into this program. By using an ROC analysis with the k-fold cross-validation method, we obtained an Az value (area under the ROC curve). For the k-fold cross-validation method, the 17 734 ROIs in the data set were randomly divided into k groups. The first group was set aside, and the remaining (k – 1) groups were used to train the ANN. The training procedure was stopped when the improvement of error distortion was less than 0.01. Moreover, the maximal number of iterations was limited to 10 000. Once trained, the network was then tested with the group that was set aside. The second group was then removed, the remaining (k – 1) groups were trained, and the network was tested with the excluded group. This process was repeated until all k groups were used, in turn, as each group was set aside and used for testing. In the simulations, k was five, and each group had 3546 or 3547 ROIs.

In addition to the ROI-based analysis described earlier, a lesion-based analysis was performed. Lesion-based analysis may be a more practically meaningful approach to evaluating the performance of the CAD system in facilitating detection of GGO. A group of ROIs that was identified as positive for GGO by using the CAD system and that was contiguous was considered as one positive cluster. If a positive cluster overlapped at least one section that contained positive ROIs (GGO ROIs), which were defined by using the reference standard in each of the 29 identified regions of GGO, the cluster was considered a true-positive cluster. If a positive cluster did not overlap any of the GGO ROIs, the cluster was considered a false-positive cluster. If one of the identified GGO regions had no overlapping positive cluster, it was considered a false-negative region.

Comparison of the Threshold Method and the Gaussian Curve Fitting Method
An ROC analysis was conducted to compare the performance of the threshold method with that of the gaussian curve fitting method. For the threshold method, the threshold was set at a range of –300 to –600 HU. The proportion of the pixels with values between –300 and –600 HU in a 30 x 30 ROI was calculated to normalize the data between 0 and 1.0. Nonparametric ROC analysis was performed by using a statistical software program (MedCalc; MedCalc Software, Mariakerke, Belgium). This analysis, which was based on the method developed by Hanley and McNeil (25), yields an empirical ROC curve and nonparametric estimate of Az with a 95% confidence interval. Two paired ROC curves were compared by means of a z test, as described by Hanley and McNeil (26).


   Results
TOP
 ABSTRACT
 INTRODUCTION
 Materials and Methods
 Results
Discussion
 Appendix
 References
 
Figure 1 illustrates the ROC curve of the ANN for classification of GGO ROIs and non-GGO ROIs. The performance of the ANN was evaluated by examining the ROC area index Az with respect to the output values that were acquired during testing. Our method had an Az value of 0.92.



View larger version (15K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 1. Graph depicts ROC curve of neural network for classification of GGO ROIs and non-GGO ROIs. Az = 0.92.

 
For statistical calculations, a threshold of 0.9 was used, and the accuracy, sensitivity, specificity, positive predictive value, and negative predictive value of the ANN for the detection of GGO ROIs by using the ROI-based analysis were 96.1% (17 034 of 17 734 ROIs), 94.3% (280 of 297 ROIs), 96.1% (16 754 of 17 437 ROIs), 29.1% (280 of 963 ROIs), and 99.9% (16 754 of 16 771 ROIs), respectively.

For the lesion-based analysis, 26 of 29 regions of GGO demonstrated true-positive findings, and three demonstrated false-negative findings. The total number of false-positive clusters was 65, which corresponds to 0.89 false-positive clusters per section. This apparent discrepancy between the relatively small number of false-positive clusters and the relatively large number of false-positive ROIs was caused by the large number (more than 40 ROIs) of false-positive clusters in one patient who had areas of GGO that spanned seven CT sections. For illustration, examples of true-positive, true-negative, false-positive, and false-negative findings from the lesion-based analysis are shown in Figure 2.



View larger version (103K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 2a. CT scans obtained with threshold of 0.9 demonstrate true-positive, true-negative, false-positive, and false-negative findings. (a) Transverse CT scan in 69-year-old woman shows true-positive finding (arrow) in right middle lobe. Although dependent opacity (arrowheads) can be categorized as GGO according to attenuation, these areas were correctly classified as negative (true-negative finding). (b) Transverse CT scan in 69-year-old woman shows localized GGO (arrow) in right upper lobe that was not detected (false-negative finding). (c) Transverse CT scan in 73-year-old woman shows false-positive finding (arrow) near airways and vessel. Arrowhead indicates true-positive finding. (d) Transverse CT scan in 61-year-old woman shows true-positive finding (arrow) in left upper lobe. Arrowheads indicate false-positive finding that was caused by artifact.

 


View larger version (78K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 2b. CT scans obtained with threshold of 0.9 demonstrate true-positive, true-negative, false-positive, and false-negative findings. (a) Transverse CT scan in 69-year-old woman shows true-positive finding (arrow) in right middle lobe. Although dependent opacity (arrowheads) can be categorized as GGO according to attenuation, these areas were correctly classified as negative (true-negative finding). (b) Transverse CT scan in 69-year-old woman shows localized GGO (arrow) in right upper lobe that was not detected (false-negative finding). (c) Transverse CT scan in 73-year-old woman shows false-positive finding (arrow) near airways and vessel. Arrowhead indicates true-positive finding. (d) Transverse CT scan in 61-year-old woman shows true-positive finding (arrow) in left upper lobe. Arrowheads indicate false-positive finding that was caused by artifact.

 


View larger version (110K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 2c. CT scans obtained with threshold of 0.9 demonstrate true-positive, true-negative, false-positive, and false-negative findings. (a) Transverse CT scan in 69-year-old woman shows true-positive finding (arrow) in right middle lobe. Although dependent opacity (arrowheads) can be categorized as GGO according to attenuation, these areas were correctly classified as negative (true-negative finding). (b) Transverse CT scan in 69-year-old woman shows localized GGO (arrow) in right upper lobe that was not detected (false-negative finding). (c) Transverse CT scan in 73-year-old woman shows false-positive finding (arrow) near airways and vessel. Arrowhead indicates true-positive finding. (d) Transverse CT scan in 61-year-old woman shows true-positive finding (arrow) in left upper lobe. Arrowheads indicate false-positive finding that was caused by artifact.

 


View larger version (75K):
[in this window]
[in a new window]
[Download PPT slide]
  		Figure 2d. CT scans obtained with threshold of 0.9 demonstrate true-positive, true-negative, false-positive, and false-negative findings. (a) Transverse CT scan in 69-year-old woman shows true-positive finding (arrow) in right middle lobe. Although dependent opacity (arrowheads) can be categorized as GGO according to attenuation, these areas were correctly classified as negative (true-negative finding). (b) Transverse CT scan in 69-year-old woman shows localized GGO (arrow) in right upper lobe that was not detected (false-negative finding). (c) Transverse CT scan in 73-year-old woman shows false-positive finding (arrow) near airways and vessel. Arrowhead indicates true-positive finding. (d) Transverse CT scan in 61-year-old woman shows true-positive finding (arrow) in left upper lobe. Arrowheads indicate false-positive finding that was caused by artifact.

 
The Az value was 0.849 for the threshold method and 0.949 for the gaussian curve fitting method. The 95% confidence intervals were 0.843 and 0.854 for the threshold method and 0.945 and 0.952 for the gaussian curve fitting method. A comparison of Az values for GGO detection demonstrated a significant difference (P < .001) between the threshold method and the gaussian curve fitting method.


   Discussion
TOP
 ABSTRACT
 INTRODUCTION
 Materials and Methods
 Results
 Discussion
Appendix
 References
 
GGO is a nonspecific finding that may be caused by various disorders, including inflammatory disease, fibrosis, and neoplastic disease (27–31). GGO is associated not only with pathologic findings but also with physiologic conditions (32). Any condition that decreases the air content of the lung parenchyma without completely obliterating the alveoli can produce GGO. Localized GGO that is greater than or equal to 1 cm in diameter and GGO that has a solid component are important clinical signs of malignancy (8,12). Results of some studies suggest that the patients who have GGO components that occupied more than 50% of a malignant nodule have a better prognosis than those who have GGO components that occupied less than 50% of a malignant nodule (10,33,34). Therefore, the detection of localized GGO is an important task in lung cancer screening.

Defining GGO on the basis of attenuation value alone is incomplete and subjective. Simply setting an attenuation range for GGO is inadequate because other structures in the lung would also be visualized in that same attenuation range. GGO may be quantified with texture analysis, which can then be used to differentiate GGO from various structures that have soft-tissue attenuation values. Texture-based analysis has been applied to classify parenchyma patterns in CT data sets (16,17,35,36). Data from previous studies (15,18) indicate that the application of a "density mask" (attenuation range, –750 to –300 HU for GGO) without texture features, which are used to facilitate detection of diffuse GGO at high-resolution CT, results in a high number of false-positive findings at air-tissue interfaces, such as the bronchovascular bundle and pleura. In our study, the mean attenuation of the localized GGO that we evaluated ranged from –740 to –174 HU (mean, –435 HU). We also found that the gaussian curve fitting method was superior to the fixed threshold method for facilitating detection of GGO. The gaussian curve fitting operation of the histogram that was used in our study improved detection accuracy by providing a more dynamic and discriminative range of attenuation values for the detection of GGO than was provided by the commonly used fixed range of attenuation values.

The selection of texture features will affect the diagnostic performance of the final CAD scheme. Optimal selection of an adequate set of texture features is a challenge, especially with the limited data obtained in our study. The larger the number of meaningful features used, the better the classification of the sample. If too many features are used, however, the ability to generalize results to larger populations may actually decrease owing to the overfitting of samples.

In the current scheme of localized GGO detection, the performance of the ANN, as measured by using Az values, was 0.92, and the accuracy was 96.1% with a threshold of 0.9. There was a relatively large number of false-positive findings, which explains the relatively low positive predictive value. Structures that were classified as false-positive findings included partial-volume averaging of the dome of the diaphragm, beam-hardening artifacts caused by dense contrast material in large veins, and interlacing pulmonary vessels and airways. In practice, these structures should be readily recognized as false-positive findings because of their characteristic locations. Conversely, most cardiac or respiratory motion artifacts and dependent opacities were correctly classified as non-GGO and were therefore not included in the list of false-positive findings. Some regions of GGO that were small or had high attenuation values were not detected and constituted false-negative findings.

There are several limitations to this study. First, the number of cases in our study was small. This small sample size reflects the fact that localized GGO is a less common manifestation of pulmonary malignancy than solid nodules. This limitation was somewhat inevitable at this stage of development, and a further study should be conducted on a larger scale to validate these preliminary results. Second, because of the retrospective nature of data collection, the CT scanning parameters were variable. The effects of scanning parameters, such as section thickness and reconstruction algorithms, could be investigated in future studies. Third, no histopathologic proof was obtained for some cases of localized GGO. When there were multiple regions of GGO, as in cases of primary lung neoplasm, only one lesion biopsy was performed and used for diagnosis. No biopsies were performed on GGO lesions for a presumed pulmonary metastasis. This limitation may not be detrimental to our study because we were focused mainly on the detection rather than the characterization or diagnostic evaluation of lesions. Fourth, although most false-positive findings were easily identified by observers, the high number of false-positive findings could become a limitation, particularly for thin-collimation multi–detector row CT scanning of the entire lungs.

In conclusion, a computerized scheme that is based on the application of ANNs to selected texture features and gaussian curve fitting features may hold promise for facilitating detection of localized GGO at CT.


   Appendix
TOP
 ABSTRACT
 INTRODUCTION
 Materials and Methods
 Results
 Discussion
 Appendix
References
 
A gaussian model was fitted to the attenuation histogram of the pixels within each ROI as G(x) = A · exp[–(x – xc)2/w2], where G(x) represents the gaussian function with A as the height of the curve, xc is the center of the curve, and w is the width of the curve. The area under the fitted curve (AUC) within the range (a, b) can be calculated as follows:

where

When the area occupied by localized GGO within an ROI is smaller than the area occupied by the background pulmonary parenchyma (as frequently occurs), the segment of the histogram that corresponds to the localized GGO may be too small to fit a gaussian curve without additional constraints. Because the attenuation value of localized GGO is higher (or less negative) than that of background pulmonary parenchyma, we applied a constraint so that xc would be between –300 and –600 HU for GGO. The range (a, b) is also fixed between –300 and –600 HU. The gaussian curve fitting was optimized by using the Lovenberg-Marquardt algorithm (37), which was modified for this constraint. From the fitted curves for GGO in ROIs, areas under the curve for GGO were calculated.


   FOOTNOTES
 

Abbreviations: ANN = artificial neural network • Az = area under ROC curve • CAD = computer-aided diagnosis • GGO = ground-glass opacity • ROC = receiver operating characteristic • ROI = region of interest

Authors stated no financial relationship to disclose.

Author contributions: Guarantors of integrity of entire study, K.G.K., J.M.G., K.T.B.; study concepts/study design or data acquisition or data analysis/interpretation, all authors; manuscript drafting or manuscript revision for important intellectual content, all authors; approval of final version of submitted manuscript, all authors; literature research, K.G.K., J.M.G., J.H.K., H.J.L.; clinical studies, J.M.G., H.J.L., J.G.I.; statistical analysis, K.G.K., J.M.G., B.G.M.; and manuscript editing, J.M.G., K.T.B.


   References
TOP
 ABSTRACT
 INTRODUCTION
 Materials and Methods
 Results
 Discussion
 Appendix
 References
 

  1. 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]
  2. Armato SG III, Li F, Giger ML, MacMahon H, Sone S, Doi K. Lung cancer: performance of automated lung nodule detection applied to cancers missed in a CT screening program. Radiology 2002;225:685–692.[Abstract/Free Full Text]
  3. 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]
  4. 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]
  5. Goo JM, Lee JW, Lee HJ, Kim S, Kim JH, Im JG. Automated lung nodule detection at low-dose CT: preliminary experience. Korean J Radiol 2003;4:211–216.[Medline]
  6. Awai K, Murao K, Ozawa A, et al. Pulmonary nodules at chest CT: effect of computer-aided diagnosis on radiologists' detection performance. Radiology 2004;230:347–352.[Abstract/Free Full Text]
  7. Lee JW, Goo JM, Lee HJ, Kim JH, Kim S, Kim YT. The potential contribution of a computer-aided detection system for lung nodule detection in multidetector row computed tomography. Invest Radiol 2004;39:649–655.[CrossRef][Medline]
  8. Henschke CI, Yankelevitz DF, Mirtcheva R, et al. CT screening for lung cancer: frequency and significance of part-solid and nonsolid nodules. AJR Am J Roentgenol 2002;178:1053–1057.[Abstract/Free Full Text]
  9. Jang HJ, Lee KS, Kwon OJ, Rhee CH, Shim YM, Han J. Bronchioloalveolar carcinoma: focal area of ground-glass attenuation at thin-section CT as an early sign. Radiology 1996;199:485–488.[Abstract/Free Full Text]
  10. Aoki T, Nakata H, Watanabe H, et al. Evolution of peripheral lung adenocarcinomas: CT findings correlated with histology and tumor doubling time. AJR Am J Roentgenol 2000;174:763–768.[Abstract/Free Full Text]
  11. Nakajima R, Yokose T, Kakinuma R, Nagai K, Nishiwaki Y, Ochiai A. Localized pure ground-glass opacity on high-resolution CT: histologic characteristics. J Comput Assist Tomogr 2002;26:323–329.[CrossRef][Medline]
  12. Nakata M, Saeki H, Takata I, et al. Focal ground-glass opacity detected by low-dose helical CT. Chest 2002;121:1464–1467.[Abstract/Free Full Text]
  13. Kakinuma R, Ohmatsu H, Kaneko M, et al. Progression of focal pure ground-glass opacity detected by low-dose helical computed tomography screening for lung cancer. J Comput Assist Tomogr 2004;28:17–23.[CrossRef][Medline]
  14. Li F, Sone S, Abe H, MacMahon H, Armato SG III, Doi K. Lung cancers missed at low-dose helical CT screening in a general population: comparison of clinical, histopathologic, and imaging findings. Radiology 2002;225:673–683.[Abstract/Free Full Text]
  15. Heitmann KR, Kauczor H, Mildenberger P, Uthmann T, Perl J, Thelen M. Automatic detection of ground glass opacities on lung HRCT using multiple neural networks. Eur Radiol 1997;7:1463–1472.[CrossRef][Medline]
  16. Uppaluri R, Hoffman EA, Sonka M, Hartley PG, Hunninghake GW, McLennan G. Computer recognition of regional lung disease patterns. Am J Respir Crit Care Med 1999;160:648–654.[Abstract/Free Full Text]
  17. Uchiyama Y, Katsuragawa S, Abe H, et al. Quantitative computerized analysis of diffuse lung disease in high-resolution computed tomography. Med Phys 2003;30:2440–2454.[CrossRef][Medline]
  18. Kauczor HU, Heitmann K, Heussel CP, Marwede D, Uthmann T, Thelen M. Automatic detection and quantification of ground-glass opacities on high-resolution CT using multiple neural networks: comparison with a density mask. AJR Am J Roentgenol 2000;175:1329–1334.[Abstract/Free Full Text]
  19. 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]
  20. Haralick RM. Statistical and structural approaches to texture. Proc IEEE 1979;67:786–804.
  21. Parker JR. Algorithms for image processing and computer vision. New York, NY: Wiley, 1997;150–175.
  22. Pitas I. Digital image processing algorithms and applications. 2nd ed. Hoboken, NJ: Wiley-Interscience, 2000;275–321.
  23. Metz CE. ROC methodology in radiologic imaging. Invest Radiol 1986;21:720–733.[Medline]
  24. Park SH, Goo JM, Jo CH. Receiver operating characteristic (ROC) curve: practical review for radiologists. Korean J Radiol 2004;5:11–18.[Medline]
  25. Hanley JA, McNeil BJ. The meaning and use of the area under a receiver operating characteristic (ROC) curve. Radiology 1982;143:29–36.[Abstract/Free Full Text]
  26. Hanley JA, McNeil BJ. A method comparing the areas under receiver operator characteristic curves derived from the same cases. Radiology 1983;148:839–843.[Abstract/Free Full Text]
  27. Nowers K, Rasband JD, Berges G, Gosselin M. Approach to ground-glass opacification of the lung. Semin Ultrasound CT MR 2002;23:302–323.[CrossRef][Medline]
  28. Kim Y, Lee KS, Jung KJ, Han J, Kim JS, Suh JS. Halo sign on high resolution CT: findings in spectrum of pulmonary diseases with pathologic correlation. J Comput Assist Tomogr 1999;23:622–626.[CrossRef][Medline]
  29. Moon JH, Kim EA, Lee KS, Kim TS, Jung KJ, Song JH. Cytomegalovirus pneumonia: high-resolution CT findings in ten non-AIDS immunocompromised patients. Korean J Radiol 2000;1:73–78.[Medline]
  30. Kim SJ, Lee KS, Ryu YH, et al. Reversed halo sign on high-resolution CT of cryptogenic organizing pneumonia: diagnostic implications. AJR Am J Roentgenol 2003;180:1251–1254.[Abstract/Free Full Text]
  31. Han D, Goo JM, Im JG, Lee KS, Paek DM, Park SH. Thin-section CT findings of arc-welders' pneumoconiosis. Korean J Radiol 2000;1:79–83.[Medline]
  32. Lee KN, Yoon SK, Sohn CH, Choi PJ, Webb WR. Dependent lung opacity at thin-section CT: evaluation by spirometrically-gated CT of the influence of lung volume. Korean J Radiol 2002;3:24–29.[Medline]
  33. Kodama K, Higashiyama M, Yokouchi H, et al. Prognostic value of ground-glass opacity found in small lung adenocarcinoma on high-resolution CT scanning. Lung Cancer 2001;33:17–25.[CrossRef][Medline]
  34. Aoki T, Tomoda Y, Watanabe H, et al. Peripheral lung adenocarcinoma: correlation of thin-section CT findings with histologic prognostic factors and survival. Radiology 2001;220:803–809.[Abstract/Free Full Text]
  35. Chabat F, Yang GZ, Hansell DM. Obstructive lung diseases: texture classification for differentiation at CT. Radiology 2003;228:871–877.[Abstract/Free Full Text]
  36. Hoffman EA, Reinhardt JM, Sonka M, et al. Characterization of the interstitial lung diseases via density-based and texture-based analysis of computed tomography images of lung structure and function. Acad Radiol 2003;10:1104–1118.[CrossRef][Medline]
  37. Marquardt DW. An algorithm for least-squares estimations of nonlinear parameters. J Soc Indust Appl Math 1963;11:431–441.[CrossRef]



This article has been cited by other articles:


Home page
 ChestHome page
C. M. Park, J. M. Goo, T. J. Kim, H. J. Lee, K. W. Lee, C. H. Lee, Y. T. Kim, K. G. Kim, H. Y. Lee, E.-A. Park, et al.
Pulmonary Nodular Ground-Glass Opacities in Patients With Extrapulmonary Cancers: What is Their Clinical Significance and How Can We Determine Whether They Are Malignant or Benign Lesions?
Chest, June 1, 2008; 133(6): 1402 - 1409.
[Abstract] [Full Text] [PDF]

Home page
 RadioGraphicsHome page
C. M. Park, J. M. Goo, H. J. Lee, C. H. Lee, E. J. Chun, and J.-G. Im
Nodular Ground-Glass Opacity at Thin-Section CT: Histologic Correlation and Evaluation of Change at Follow-up
RadioGraphics, March 1, 2007; 27(2): 391 - 408.
[Abstract] [Full Text] [PDF]

This Article
Right arrow Abstract Freely available
Right arrow Figures Only
Right arrow Full Text (PDF)
Right arrow All Versions of this Article:
2372041461v1
237/2/657    most recent
Right arrow Submit a response
Right arrow Alert me when this article is cited
Right arrow Alert me when eLetters are posted
Right arrow Alert me if a correction is posted
Right arrow Citation Map
Services
Right arrow Email this article to a friend
Right arrow Similar articles in this journal
Right arrow Similar articles in PubMed
Right arrow Alert me to new issues of the journal
Right arrow Download to citation manager
Right arrow reprints & permissions
Citing Articles
Right arrow Citing Articles via HighWire
Right arrow Citing Articles via Google Scholar
Google Scholar
Right arrow Articles by Kim, K. G.
Right arrow Articles by Im, J.-G.
Right arrow Search for Related Content
PubMed
Right arrow PubMed Citation
Right arrow Articles by Kim, K. G.
Right arrow Articles by Im, J.-G.

HOME HELP FEEDBACK SUBSCRIPTIONS ARCHIVE SEARCH TABLE OF CONTENTS
RADIOLOGY RADIOGRAPHICS RSNA JOURNALS ONLINE