computerized texture analysis of persistent part-solid ground-glass nodules: differentiation of...

Radiology: Volume 273: Number 1—October 2014 n radiology.rsna.org 285

Original research n

Thoracic imaging

1 From the Department of Radiology, Seoul National University College of Medicine, and Institute of Radiation Medicine, Seoul National University Medical Research Center, 101 Daehangno, Jongno-gu, Seoul 110-744, Korea (H.D.C., C.M.P., S.J.P., S.M.L., J.M.G.); Cancer Research Institute, Seoul National University, Seoul, Korea (C.M.P., S.J.P., J.M.G.); and Department of Biomedical Engineering, Division of Basic & Applied Sciences, National Cancer Cen-ter, Gyeonggi-Do, Korea (K.G.K.). Received September 25, 2013; revision requested November 7; revision received April 18, 2014; accepted May 13; final version accepted June 11. Supported by a Research Grant from the Korean Foundation for Cancer Research (grant CB-2011-02-01) and Basic Science Research Program through the National Research Foundation of Korea funded by the Ministry of Education, Science and Technology (grant 2011-0022379). Address correspondence to C.M.P. (e-mail: [email protected]).

q RSNA, 2014

Purpose: To retrospectively investigate the value of computerized three-dimensional texture analysis for differentiation of preinvasive lesions from invasive pulmonary adenocar-cinomas (IPAs) that manifest as part-solid ground-glass nodules (GGNs).

Materials and Methods:

The institutional review board approved this retrospec-tive study with a waiver of patients’ informed consent. The study consisted of 86 patients with 86 pathologic analysis–confirmed part-solid GGNs (mean size, 16 mm 6 5.4 [standard deviation]) who had undergone computed tomographic (CT) imaging between January 2005 and Oc-tober 2011. Each part-solid GGN was manually segmented and its computerized texture features were quantitatively extracted by using an in-house software program. Mul-tivariate logistic regression analysis was performed to investigate the differentiating factors of preinvasive le-sions from IPAs. Three-layered artificial neural networks (ANNs) with a back-propagation algorithm and receiver operating characteristic curve analysis were used to build a discriminating model with texture features and to evalu-ate its discriminating performance.

Results: Pathologic analysis confirmed 58 IPAs (seven minimally invasive adenocarcinomas and 51 invasive adenocarci-nomas) and 28 preinvasive lesions (four atypical adeno-matous hyperplasias and 24 adenocarcinomas in situ). IPAs and preinvasive lesions exhibited significant differ-ences in various histograms and volumetric parameters (P , .05). Multivariate analysis revealed that smaller mass (adjusted odds ratio, 0.092) and higher kurtosis (adjusted odds ratio, 3.319) are significant differentiators of preinvasive lesions from IPAs (P , .05). With mean attenuation, standard deviation of attenuation, mass, kur-tosis, and entropy, the ANNs model showed excellent ac-curacy in differentiation of preinvasive lesions from IPAs (area under the curve, 0.981).

Conclusion: In part-solid GGNs, higher kurtosis and smaller mass are significant differentiators of preinvasive lesions from IPAs, and preinvasive lesions can be accurately differentiated from IPAs by using computerized texture analysis.

q RSNA, 2014

Online supplemental material is available for this article.

Hee-Dong Chae, MDChang Min Park, MDSang Joon Park, PhDSang Min Lee, MDKwang Gi Kim, PhDJin Mo Goo, MD

computerized Texture analysis of Persistent Part-solid ground-glass nodules: Differentiation of Preinvasive Lesions from Invasive Pulmonary Adenocarcinomas1

Note: This copy is for your personal non-commercial use only. To order presentation-ready copies for distribution to your colleagues or clients, contact us at www.rsna.org/rsnarights.

THORACIC IMAGING: Computerized Texture Analysis of Persistent Part-Solid Ground-Glass Nodules Chae et al

286 radiology.rsna.org n Radiology: Volume 273: Number 1—October 2014

The purpose of our study was to retrospectively investigate the value of computerized three-dimensional tex-ture analysis for differentiation of pre-invasive lesions from IPAs that manifest as part-solid GGNs.

Materials and Methods

Our institutional review board approved this retrospective study with a waiver of informed consent.

Study PopulationOne author (S.M.L.) searched the electronic medical records and the radiology information systems of our hospital between January 2005 and October 2011 and selected all CT ex-aminations with the descriptive terms “GGO,” “GGN,” “part-solid nodule,” “nonsolid nodule,” “subsolid nodule,” “ground-glass opacity,” or “ground-glass nodule.” A total of 16 720 CT examinations in 9942 patients were identified, out of which 4010 patients with 7759 CT examinations had avail-able CT images with a section thick-ness of 2.5 mm or less. Thereafter, diffuse ground-glass opacity, very small GGNs (,5 mm), large lesions (.3 cm), and transient GGNs were

the solid component is larger than 5 mm, should be considered malignant until proven otherwise and should be considered candidates for surgical re-sections. However, more than a third of part-solid GGNs are preinvasive le-sions (4), and there is evidence that these preinvasive lesions can be man-aged with close follow-up alone or safely treated with limited resection even if they eventually have to be re-sected (6–8). Therefore, it is impor-tant to accurately differentiate these preinvasive lesions from invasive pul-monary adenocarcinomas (IPA) that appear as part-solid GGNs. Until now, it was a challenge to differentiate these two groups through visual assessment of morphologic structure based on CT imaging because there are consider-able overlaps among AAH, AIS, and IPA (5,9–11).

In this context, computerized tex-ture analysis can be a promising tool for lesion characterization and its dif-ferentiation. Texture analysis is a quan-titative imaging analysis tool that uses attenuation values of each voxel and their distribution within target lesions, which is expected to enable a more de-tailed and reproducible quantitative as-sessment of lesion characteristics than visual analysis by human observers. Indeed, several previous studies have previously evaluated GGNs by using a texture feature such as histographic analysis (12–15). However, the lim-itation of those studies is that they were performed with two-dimensional analysis that used only histographic features (12). In addition, they had not focused on the differentiation between preinvasive lesions and invasive adeno-carcinomas, but rather on the differ-entiation of premalignant lesions (ie, AAH) from malignant lesions, includ-ing AIS (ie, former bronchiolo-alveolar carcinoma) and IPA (13).

Published online before print10.1148/radiol.14132187 Content codes:

Radiology 2014; 273:285–293

Abbreviations:AAH = atypical adenomatous hyperplasiaAIS = adenocarcinoma in situANN = artificial neural networkGGN = ground-glass noduleIPA = invasive pulmonary adenocarcinoma

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

Conflicts of interest are listed at the end of this article.

See also the article by Song et al in this issue.

Advances in Knowledge

n Invasive pulmonary adenocarci-nomas (IPAs) and preinvasive lesions that appear as part-solid ground-glass nodules (GGNs) showed significant differences in histograms and with volumetric parameters such as mass, kurto-sis, and entropy (P , .001).

n Smaller mass (adjusted odds ratio, 0.092) and higher kurtosis (adjusted odds ratio, 3.319) were significant differentiators of pre-invasive lesions from IPAs that appear as part-solid GGNs (P , .05).

n When mean attenuation, stan-dard deviation of attenuation, mass, kurtosis, and entropy were used as input variables in an arti-ficial neural network model to differentiate preinvasive lesions from IPAs that appear as part-solid GGNs, good accuracy was observed (area under the curve, 0.981).

Implication for Patient Care

n Computerized texture analysis can accurately differentiate pre-invasive lesions from invasive pulmonary adenocarcinomas that appear as part-solid GGNs.

In 2011, a new classification of pul-monary adenocarcinoma was intro-duced by the joint working group,

which consisted of the International Association for the Study of Lung Can-cer, American Thoracic Society, and European Respiratory Society (1). A major change in this classification was the introduction of adenocarcinoma in situ (AIS) as a second preinvasive le-sion of lung adenocarcinoma in addition to atypical adenomatous hyperplasia (AAH) (1). AIS is defined as a localized adenocarcinoma 3 cm or smaller that exhibits a lepidic growth pattern of neo-plastic cells along alveolar structures without stromal, vascular, or pleural in-vasion (1). AAH and AIS typically pre-sent as a persistent pure or part-solid ground-glass nodule (GGN) with small solid parts when viewed with computed tomographic (CT) imaging (2,3).

Part-solid GGNs show a greater likelihood of malignancy than do pure GGNs or solid nodules (4,5). Thus, the newly proposed Fleischner Society guideline (4) recommends that part-solid GGNs, especially those in which



classified the lung lesions as AAH, AIS, minimally invasive adenocarcinoma, or invasive adenocarcinoma accord-ing to the new lung adenocarcinoma classification (1). In our study, among the 86 pathologic analysis–proven part-solid GGNs, 28 were preinvasive lesions and 58 were IPAs. Among the 28 preinvasive lesions, four were AAH and the remaining 24 were AIS. Of the 58 IPAs, seven were minimally invasive adenocarcinomas and the re-maining 51 were invasive adenocarci-nomas (Table 1). The study population was extracted from our previous study, which investigated the differentia-tion of CT features between IPAs and preinvasive lesions that appeared as GGNs through visual assessment (16).

CT ExaminationsAll CT examinations were performed without intravenous contrast material injection by using one of three scanners (Somatom Definition, Sensation-16, Siemens Medical Solutions, Forchheim, Germany; Brilliance-64, Phillips Med-ical Systems, Best, the Netherlands; LightSpeed Ultra, LightSpeed VCT, GE Medical Systems, Waukesha, Wis). All CT examinations were performed by using dose modulation with the following parameters: 120 kVp; 100–200 mAs; pitch, 0.75–1.5; and collima-tion, 0.625–1.25 mm. All imaging data were reconstructed by using a medium-sharp reconstruction algorithm with a thickness of 1–1.25 mm. In all patients, CT images were acquired in the supine

included in this study. Subsequently, an additional group of 112 GGNs were excluded because they had no avail-able CT images without contrast en-hancement, and another 10 GGNs were excluded because they lacked CT images with a section thickness of 1.25 mm or less. Finally, our study population consisted of 86 patients (mean age, 58.8 years 6 9.4 [stan-dard deviation]; age range, 35–78 years) with 86 part-solid GGNs proved with pathologic analysis (Fig 1). There were 28 men (mean age, 60.0 years 6 9.1; age range, 36–80 years) and 58 women (mean age, 59.0 years 6 9.4; age range, 29–77 years). Among the 86 patients, 70 patients underwent lobectomy, 13 patients underwent wedge resection or seg-mentectomy, and three patients un-derwent percutaneous transthoracic core-needle biopsy.

Preinvasive lesions included AAH and AIS, and they were defined as le-sions that showed no stromal, vascu-lar, or pleural invasion (1). IPAs were defined as adenocarcinomas that con-tained an invasive component (ie, min-imally invasive adenocarcinoma and invasive adenocarcinoma, including subtypes lepidic predominant, acinar predominant, or papillary predominant adenocarcinoma) (1). An experienced pathologist (Y.K.J., with 14 years of experience in pulmonary pathology) reviewed all pathologic specimens and investigated whether there was an inva-sive component within the lesion, and

excluded, which left 303 patients with 327 GGNs. Of the 327 GGNs, 55 GGNs without a diagnosis confirmed by pathologic analysis were excluded. The remaining 272 GGNs were com-posed of 64 pure GGNs and 208 part-solid GGNs, and pure GGNs were not

Figure 1

Figure 1: Flowchart shows how the study popula-tion was selected and its retrospective manner. Numbers in parentheses are numbers of patients. GGOs = ground-glass opacities.

Table 1

Patient Characteristics

Characteristic No. of Patients

Men 28Women 58Age (y)* 58.8 6 9.4 (35–78)Pathologic subtype AAH 4 AIS 24 Minimally invasive adenocarcinoma 7 Invasive adenocarcinoma 51

Note.—Except where indicated, data are numbers of patients.

* Data are mean age 6 standard deviation. Data in parentheses are age range.



thereafter, texture features were cal-culated automatically. Manual segmen-tation of lung nodules was performed by a radiologist (H.D.C., with 2 years of experience in chest CT imaging) and confirmed by one chest radiologist (C.M.P., with 14 years of experience in chest CT). Regions of interest were de-lineated around the boundary of each part-solid GGN as depicted on CT im-ages for each section. After a nodule was segmented, texture features were

Imaging Solution for Segmentation and Texture Analysis) was used for lesion segmentation with fully auto-mated quantification of texture fea-tures implemented with a dedicated C++ language (Microsoft Foundation Classes; Microsoft, Redmond, Wash) (17). Example images are shown in Figure 2. The overall procedure of this analysis scheme was composed of two major stages: first, nodule segmen-tation was conducted manually, and

position at full inspiration. When more than one CT examination was available for CT morphologic analysis, we se-lected the last CT examination before surgery or biopsy. The mean interval between CT examination and surgery or biopsy was 31 days (range, 1–132 days; median, 25 days).

Computerized Texture AnalysisOur personal computer–based in-house software program (Medical

Figure 2

Figure 2: Screenshot shows the texture analysis software program. The segmentation of part-solid GGNs was manu-ally conducted by using an in-house software program and texture features of the nodules were automatically extracted and calculated by the software program.



using software (SPSS version 18.0; SPSS, Chicago, Ill).

A three-layered, feed-forward ar-tificial neural network (ANN) with a back-propagation algorithm was used to build a discriminating model of pre-invasive lesions. Our model consisted of five input units and one output unit that corresponded to the likelihood of a preinvasive lesion. The five input units were the same texture features that were used as input data for the logistic regression model. We utilized the leave-one-out cross-validation method to train and test our ANN model. In the k-fold cross-validation method, the whole data set is divided into k subsets, and k 2 1 of these sub-sets are then combined and used for training. The remaining set is used for testing. The algorithm continues iter-atively until each fold is used exactly once for testing, and consequently, the data used for training of ANNs are also used for testing. The leave-one-out cross-validation method is an extreme form of the k-fold cross-vali-dation method, in which the number of subsets k is equal to the number of cases (86 cases in our study), and only one case is used for testing (18). A commercially available software program (NeuroSolution; NeuroDi-mension, Gainesville, Fla) was used to implement the ANN model. The dif-ferentiating performance of the ANN model was evaluated with paramet-ric receiver operating characteristic curve analysis (Rockit version 1.1 B2; C. E. Metz, University of Chicago, Ill).

Results

Texture Features of Part-Solid GGNsThere were significant differences be-tween preinvasive lesions and IPAs that appeared as part-solid GGNs in most of the histographic parameters, such as mean attenuation, standard deviation, kurtosis, entropy, and sev-eral percentile CT numbers (P , .002) (Table 2). All volumetric parameters of preinvasive lesions were significantly different from those of IPAs. Preinva-sive lesions were significantly smaller

parameters, volumetric parameters, and morphologic features. To adjust for multiple comparison problems that resulted from an increase in the type-1 error when tested repeatedly, the Bonferroni correction method was used with an adjusted P value threshold of .002 (.05 / 22). Logistic regression analysis was conducted to identify differentiating variables of preinvasive lesions from IPAs. Char-acteristics with statistical significance (P , .002) in univariate analysis and a high clinical relevance were used as input variables for multiple logistic re-gression analysis. In multiple logistic regression analysis, a backward step-wise selection mode was used with iterative entry of variables based on test results (P , .05). The removal of variables was based on likelihood ra-tio statistics with a probability of 0.10. Interobserver variability of texture features obtained with manual seg-mentation was investigated by using intraclass correlation coefficients. In-traclass correlation coefficients less than 0.40 signified poor agreement; 0.41–0.60, moderate agreement; 0.61–0.80, good agreement; and 0.81 or greater, excellent agreement. Sta-tistical analyses were performed by

calculated and extracted automatically. In addition, another chest radiologist (S.M.L., with 8 years of experience in chest CT imaging) performed manual segmentation independently to eval-uate the interobserver variability of manual segmentation in calculating texture features of part-solid GGNs. Analyzed texture features included his-togram features, volumetric features, and morphologic features. Histogram parameters analyzed included mean attenuation, standard deviation, skew-ness, kurtosis, entropy, homogeneity, and percentile CT numbers. Volumet-ric parameters included volume, mass, effective diameter, and surface area. Finally, the following morphologic features were obtained from the nod-ules: sphericity, discrete compactness, gray-level co-occurrence matrix iner-tia, gray-level co-occurrence matrix inverse difference moment, and gray-level co-occurrence matrix contrast. See Appendix E1 (online) for detailed information on texture features.

Statistical AnalysisStatistical differences between pre-invasive lesions and IPAs were ana-lyzed by using the independent sam-ple t test for differences in histogram

Table 2

Histographic Parameters and Percentile CT Numbers of Preinvasive Lesions and Invasive Pulmonary Adenocarcinomas in Part-Solid GGNs

Characteristic Preinvasive Lesions (n = 28)Invasive Pulmonary Adenocarcinomas (n = 58) P Value*

Mean attenuation (HU) 2574.3 6 132.2 2439.7 6 114.6 ,.001Standard deviation (HU) 177.7 6 47.9 233.1 6 47.9 ,.001Skewness 0.87 6 0.68 0.33 6 0.52 .001Kurtosis 1.39 6 2.22 20.28 6 0.81 .001Entropy 6.18 6 0.28 6.61 6 0.26 ,.001Homogeneity 0.0289 6 0.0039 0.0284 6 0.0026 .544Fifth percentile (HU) 2821.6 6 82.3 2786.6 6 71.4 .04610th percentile (HU) 2781.2 6 87.2 2728.5 6 79.3 .00625th percentile (HU) 2701.3 6 108.2 2617.5 6 105.0 .00150th percentile (HU) 2597.8 6 142.5 2460.6 6 141.2 ,.00175th percentile (HU) 2475.0 6 177.7 2275.1 6 156.9 ,.00190th percentile (HU) 2338.9 6 184.8 2115.7 6 138.8 ,.00195th percentile (HU) 2244.8 6 178.5 235.5 6 119.8 ,.001

Note.—Except where indicated, data are mean 6 standard deviation.

* Independent sample t test.



than IPAs in terms of volume (1349.7 mm3 6 981.4 vs 2951.8 mm3 6 2069.8, respectively; P , .001) and mass (530.6 mg 6 367.8 vs 1672.6 mg 6 1258.5, respectively; P , .001). However, most morphologic features showed no sig-nificant differences between the two groups except for sphericity and gray-level co-occurrence matrix contrast. Fifth and 10th percentile numbers (P = .046 and .006, respectively) turned out to be statistically insignificant after mul-tiple comparison correction with Bon-ferroni method. Tables 2 and 3 sum-marize the comparison of histographic parameters, volumetric parameters, and morphologic features between pre-invasive lesions and IPAs.

Logistic Regression AnalysisWe used mean attenuation, standard deviation, mass, kurtosis, and entropy as input variables for multiple logistic regression analysis. Logistic regres-sion analysis revealed that smaller mass (adjusted odds ratio, 0.092 [95% confidence interval: 0.019, 0.452]; P = .003) and higher kurtosis (adjusted odds ratio, 3.319 [95% confidence in-terval: 1.185, 9.302]; P = .022) were statistically significant independent differentiators of preinvasive lesions from IPA (Table 4).

Differentiating Performance of the ANN Model between Preinvasive Lesions and IPAsThe same features used as input var-iables for logistic regression analysis were used as input variables for the ANN model. By using these charac-teristics, the performance of the ANN model for differentiation of preinva-sive lesions from IPA showed excellent accuracy (area under the curve, 0.981 [95% confidence interval: 0.973, 0.987]; Fig 3). Representative cases are shown in Figure 4a and 4b.

Interobserver Variability AnalysisThe results of interobserver variabil-ity for texture features obtained with manual segmentation by two radiolo-gists are summarized in Table E1 (on-line). All texture features calculated from two sets of regions of interest

Figure 3

Figure 3: Graph shows the performance of the ANN model for dis-crimination of preinvasive lesions from IPA. The results of the receiver operating characteristic curve analysis of the ANN model in discriminat-ing preinvasive lesions from invasive pulmonary adenocarcinomas that appear as part-solid GGNs are shown. The area under the curve of the model was 0.981, which consisted of mean, standard deviation, mass, kurtosis, and entropy.

Table 3

Volumetric Parameters and Morphologic Features of Preinvasive Lesions and Invasive Pulmonary Adenocarcinomas in Part-Solid GGNs

Characteristic Preinvasive Lesions (n = 28) Invasive Lesions (n = 58) P Value*

Volume (mm3) 1349.7 6 981.4 2951.8 6 2069.8 ,.001Mass (mg) 530.6 6 367.8 1672.6 6 1258.5 ,.001Effective diameter (mm) 12.89 6 3.39 16.85 6 4.13 ,.001Surface area (mm2) 680.7 6 373.6 1207.8 6 640.1 ,.001Sphericity 0.724 6 0.109 0.642 6 0.082 ,.001Discrete compactness 0.447 6 0.167 0.431 6 0.151 .657GLCM inertia 2.17 6 0.18 2.11 6 0.54 .103GLCM IDM 0.0123 6 0.0034 0.0116 6 0.0020 .212GLCM contrast 2.53 6 1.42 4.11 6 1.50 ,.001

Note.—Except where indicated, data are mean 6 standard deviation. GLCM = gray-level co-occurrence matrix, IDM = inverse

difference moment.

* Independent sample t test.

Table 4

Results of Logistic Regression Analysis for CT Determination of Preinvasive Lesions and Invasive Pulmonary Adenocarcinomas in Part-Solid GGNs

Variable Adjusted Odds Ratio P Value

Mass (mg) 0.092 (0.019, 0.452) .003Kurtosis 3.319 (1.185, 9.302) .022

Note.—Data are adjusted odds ratios per one standard deviation change; data in parentheses are 95% confidence intervals.



peaks) could be useful to differenti-ate AAHs from bronchiolo-alveolar carcinomas and IPAs. However, these previous studies have only focused on discriminating AAHs from bronchiolo-alveolar carcinomas and IPAs. To our knowledge, there are no studies that attempted to discriminate preinvasive lesions from IPAs by using the com-puterized texture analysis method.

In our study, we found that mass and kurtosis were significant differ-entiators of preinvasive lesions from

quantitative methodologic analysis to evaluate GGNs. Nomori et al (12) ana-lyzed histograms of CT pixel numbers of AAHs and bronchiolo-alveolar car-cinomas by using two-dimensional section images and found that peak CT attenuation value was useful to differentiate AAHs from bronchiolo-alveolar carcinomas. Ikeda et al (13) extended this approach into three-dimensional quantification of GGNs and showed that a CT number–based histographic pattern (ie, one or two

that were manually segmented by two radiologists showed good or excellent agreement.

Discussion

The assessment of GGNs, particularly the assessment of part-solid GGNs by using visual features, is not an easy task because of substantial overlaps among imaging features of AAH, AIS, and IPA. To address these challenges, several investigators have adopted

Figure 4

Figure 4: CT images show preinvasive lesions and invasive adenocarcinomas with their texture features. (a) CT scan shows a 14-mm part-solid GGN (arrow) with a small solid portion in the superior segment of the left lower lobe in a 52-year-old woman. This nodule was confirmed as AIS at pathologic analysis. (b) CT scan shows a 15-mm part-solid GGN (arrow) with a lobulated border in the anterior segment of the left upper lobe in a 69-year-old man. This nodule was confirmed as adenocarcinoma at lobectomy. The preinvasive lesion AIS had a smaller mass (652.28 mg) than the invasive adenocarcinoma (857.41 mg). (c, d) The CT number distribution of the AIS had a higher peak, which meant that it had higher kurtosis (0.031). However, the adenocarcinoma had a negative kurtosis (20.469) with a lower and wider peak on the histogram.



standard for lesion segmentation, particularly in the case of part-solid GGNs because their margins are often indistinct from the normal lung paren-chyma and, consequently, not techni-cally easy to segment automatically. Nevertheless, we believe that a reli-able and robust automatic boundary extraction method should be further developed to address the variability issue. Finally, the performance of our ANN model may be overly optimis-tic because of the lack of a separate validation set and the small number of cases relative to the large number of potential predictors considered. In our study, we utilized the leave-one-out cross-validation method to over-come this limitation and to improve the generalizability of the ANN model.

In conclusion, in part-solid GGNs, higher kurtosis and smaller mass are significant differentiators of preinva-sive lesions from IPAs, and preinvasive lesions can be accurately differenti-ated from IPAs by using computerized texture analysis.

Acknowledgment: We thank Yoon Kyung Jeon, MD, Department of Pathology, Seoul National University Hospital, for reviewing pathologic specimens.

Disclosures of Conflicts of Interest: H.D.C. dis-closed no relevant relationships. C.M.P. disclosed no relevant relationships. S.J.P. disclosed no rele-vant relationships. S.M.L. disclosed no relevant re-lationships. K.G.K. disclosed no relevant relation-ships. J.M.G. disclosed no relevant relationships.

References 1. Travis WD, Brambilla E, Noguchi M, et

al. International Association for the Study of Lung Cancer/American Thoracic Soci-ety/European Respiratory Society inter-national multidisciplinary classification of lung adenocarcinoma. J Thorac Oncol 2011;6(2):244–285.

2. Park CM, Goo JM, Lee HJ, et al. CT find-ings of atypical adenomatous hyperplasia in the lung. Korean J Radiol 2006;7(2):80–86.

3. Nagao M, Murase K, Yasuhara Y, et al. Measurement of localized ground-glass at-tenuation on thin-section computed tomog-raphy images: correlation with the progres-sion of bronchioloalveolar carcinoma of the lung. Invest Radiol 2002;37(12):692–697.

4. Naidich DP, Bankier AA, MacMahon H, et al. Recommendations for the management of subsolid pulmonary nodules detected at

reflect that ANNs may be suited to modeling complex relationships, such as those found between the invasive-ness of a pulmonary nodule and tex-ture features (18). However, ANNs are also likely to be over fitting, which means that the model may be adjusted specifically to the available data set, but perform poorly for un-seen data (23). In our study, we re-stricted the size of the network by decreasing the number of input vari-ables and used k-fold cross-validation to improve the ability of the model to generalize. Future prospective studies are anticipated to validate this excel-lent performance of the ANN model to differentiate between preinvasive lesions and IPAs that appear as part-solid GGNs.

Given the positive results of the National Lung Screening Trial (24), we can anticipate an increased use of CT screening for lung cancer and a re-sultant increase in the identification of preinvasive lesions (25). Thus, the de-velopment of a systematic method to differentiate preinvasive lesions from IPAs that appear as subsolid nodules is important given that there is con-cern regarding the overdiagnosis of lung cancer. Our study suggests use-ful quantitative nodule characteristics for the differentiation of part-solid GGNs. Further studies are warranted to show the additional value of quan-titative texture analysis to nontextural measures.

Our study has several limitations. First, our study was retrospective, which means that it is subjected to potential bias. Second, our examina-tion protocol was not standardized among patients, and CT examinations were performed by using different technical parameters. This may have resulted in the variability of CT atten-uation values with resultant bias for estimation of texture features. Third, the texture features in this study were derived from the results of manual segmentation by radiologists; these results can be significantly influenced by subjective trend or bias on the part of the observers. However, manual segmentation is the current reference

IPAs that appear as part-solid GGNs. Mass is the product of the volume and density of lesions and reflects both the lesions’ size and internal attenuation, simultaneously (19,20). Therefore, mass reflects the exact tumor burden of GGNs, and we found that preinva-sive lesions had a significantly smaller volume and mass than IPAs, which can be interpreted as preinvasive le-sions having a smaller tumor burden than IPAs. AAHs or AISs are typically smaller in size and have fainter in-ternal attenuation or a smaller solid component within GGNs compared with IPAs. In fact, a previous study (16) on the visual assessment of part-solid GGNs revealed that preinvasive lesions were able to be distinguished from IPAs with smaller lesion size and smaller solid proportion. Many previ-ous studies have reported that part-solid GGNs with a greater extent of the solid component have a higher likelihood of being invasive adenocar-cinomas with an associated poorer prognosis (4,21,22). We believe that mass is more suitable for the evalua-tion of part-solid GGNs than volume or diameter of the nodule because mass can simultaneously reflect both the lesions’ size and internal attenua-tion, as described, and furthermore, it was reported to show better measure-ment reproducibility (19). Interest-ingly, preinvasive lesions had higher kurtosis compared with IPAs. Normal distribution has a kurtosis of 0 and, if the kurtosis value is greater than 0, its distribution is peaked and displays a tall and skinny shape. Unlike IPAs, which usually show a two-peak pat-tern on CT attenuation histograms, preinvasive lesions (ie, AAH and AIS) frequently show a one-peak pattern on histogram analysis (9,13). Our re-sults are in good correspondence with those of previous studies as a histo-gram with a one-peak pattern will show a more peaked appearance and have higher kurtosis than that with a two-peak pattern.

We found that ANN showed ex-cellent performance in differentiating preinvasive lesions from IPAs (area under the curve = 0.981). The results



CT: a statement from the Fleischner Soci-ety. Radiology 2013;266(1):304–317.

5. Lee HJ, Goo JM, Lee CH, Yoo CG, Kim YT, Im JG. Nodular ground-glass opacities on thin-section CT: size change during fol-low-up and pathological results. Korean J Radiol 2007;8(1):22–31.

6. Van Schil PE, Asamura H, Rusch VW, et al. Surgical implications of the new IASLC/ATS/ERS adenocarcinoma classification. Eur Respir J 2012;39(2):478–486.

7. Watanabe S, Watanabe T, Arai K, Kasai T, Haratake J, Urayama H. Results of wedge resection for focal bronchioloalveolar car-cinoma showing pure ground-glass attenua-tion on computed tomography. Ann Thorac Surg 2002;73(4):1071–1075.

8. Koike T, Togashi K, Shirato T, et al. Limited resection for noninvasive bronchioloalveo-lar carcinoma diagnosed by intraoperative pathologic examination. Ann Thorac Surg 2009;88(4):1106–1111.

9. Lee HJ, Lee CH, Jeong YJ, et al. IASLC/ATS/ERS International Multidisciplinary Classification of Lung Adenocarcinoma: novel concepts and radiologic implica-tions. J Thorac Imaging 2012;27(6):340– 353.

10. Park CM, Goo JM, Lee HJ, Lee CH, Chun EJ, Im JG. Nodular ground-glass opacity at thin-section CT: histologic correlation and evaluation of change at follow-up. Radio-Graphics 2007;27(2):391–408.

11. Kim HY, Shim YM, Lee KS, Han J, Yi CA, Kim YK. Persistent pulmonary nodular ground-glass opacity at thin-section CT: histopathologic comparisons. Radiology 2007;245(1):267–275.

12. Nomori H, Ohtsuka T, Naruke T, Suemasu K. Differentiating between atypical adeno-matous hyperplasia and bronchioloalveolar carcinoma using the computed tomogra-phy number histogram. Ann Thorac Surg 2003;76(3):867–871.

13. Ikeda K, Awai K, Mori T, Kawanaka K, Ya-mashita Y, Nomori H. Differential diagnosis

of ground-glass opacity nodules: CT number analysis by three-dimensional computerized quantification. Chest 2007;132(3):984–990.

14. Nakata M, Sawada S, Yamashita M, et al. Objective radiologic analysis of ground-glass opacity aimed at curative limited resection for small peripheral non-small cell lung can-cer. J Thorac Cardiovasc Surg 2005;129(6): 1226–1231.

15. Matsuguma H, Nakahara R, Anraku M, et al. Objective definition and measurement method of ground-glass opacity for planning limited resection in patients with clinical stage IA adenocarcinoma of the lung. Eur J Cardiothorac Surg 2004;25(6):1102–1106.

16. Lee SM, Park CM, Goo JM, Lee HJ, Wi JY, Kang CH. Invasive pulmonary ade-nocarcinomas versus preinvasive lesions appearing as ground-glass nodules: differ-entiation by using CT features. Radiology 2013;268(1):265–273.

17. Lee KH, Goo JM, Park SJ, et al. Correlation between the size of the solid component on thin-section CT and the invasive component on pathology in small lung adenocarcinomas manifesting as ground-glass nodules. J Tho-rac Oncol 2014;9(1):74–82.

18. Ayer T, Chhatwal J, Alagoz O, Kahn CE Jr, Woods RW, Burnside ES. Informatics in radiology: comparison of logistic regres-sion and artificial neural network models in breast cancer risk estimation. Radio-Graphics 2010;30(1):13–22.

19. de Hoop B, Gietema H, van de Vorst S, Murphy K, van Klaveren RJ, Prokop M. Pulmonary ground-glass nodules: increase in mass as an early indicator of growth. Ra-diology 2010;255(1):199–206.

20. Mull RT. Mass estimates by computed tomog-raphy: physical density from CT numbers. AJR Am J Roentgenol 1984;143(5):1101–1104.

21. Ohde Y, Nagai K, Yoshida J, et al. The proportion of consolidation to ground-glass opacity on high resolution CT is a good pre-dictor for distinguishing the population of

non-invasive peripheral adenocarcinoma. Lung Cancer 2003;42(3):303–310.

22. Kim EA, Johkoh T, Lee KS, et al. Quan-tification of ground-glass opacity on high-resolution CT of small peripheral ade-nocarcinoma of the lung: pathologic and prognostic implications. AJR Am J Roent-genol 2001;177(6):1417–1422.

23. Ye X, Lin X, Dehmeshki J, Slabaugh G, Bed-doe G. Shape-based computer-aided detec-tion of lung nodules in thoracic CT images. IEEE Trans Biomed Eng 2009;56(7):1810–1820.

24. National Lung Screening Trial Research Team, Aberle DR, Adams AM, et al. Re-duced lung-cancer mortality with low-dose computed tomographic screening. N Engl J Med 2011;365(5):395–409.

25. Austin JH, Garg K, Aberle D, et al. Radio-logic implications of the 2011 classification of adenocarcinoma of the lung. Radiology 2013;266(1):62–71.

26. Ganeshan B, Miles KA, Young RC, Cha-twin CR. Hepatic entropy and uniformity: additional parameters that can potentially increase the effectiveness of contrast en-hancement during abdominal CT. Clin Ra-diol 2007;62(8):761–768.

27. Sharma N, Ray AK, Sharma S, Shukla K, Aggarwal L, Pradhan S. Segmentation of medical images using simulated anneal-ing based fuzzy C Means algorithm. Int J Biomed Eng Technol 2009;2(3):260– 278.

28. Bribiesca E. An easy measure of compact-ness for 2D and 3D shapes. Pattern Recog-nit 2008;41(2):543–554.

29. Albregtsen F. Statistical texture measures computed from gray level coocurrence matrices. Image Processing Laboratory, Department of Informatics, University of Oslo Web site; http://www.uio.no/studier/emner/matnat/ifi/INF4300/h08/undervis-ningsmateriale/glcm.pdf. Published 1995. Updated November 5, 2008. Accessed Sep-tember 7, 2013.

computerized texture analysis of persistent part-solid ground-glass nodules: differentiation of...

Documents