Quantitative CT Classification of Lung Nodules: Initial Comparison of 2D and 3D Analysis

Abstract

Objective

To compare the performance of 2D and 3D quantitative CT methods for classifying lung nodules as lung cancer, metastases, or benign.

Materials and Methods

Using semiautomated software and computerized analysis, we analyzed more than 50 quantitative CT features of 96 solid nodules in 94 patients, in 2D from a single slice and in 3D from the entire nodule volume. Multivariable logistic regression was used to classify nodule types. Model performance was assessed by the area under the receiver-operating characteristic curve (AUC) using leave-one-out cross-validation.

Results

The AUC for distinguishing 53 primary lung cancers from 18 benign nodules and 25 metastases ranged from 0.79 to 0.83 and was not significantly different for 2D and 3D analyses (p=0.29–0.78). Models distinguishing metastases from benign nodules were statistically significant only by 3D analysis (AUC=0.84).

Conclusion

3D CT methods did not improve discrimination of lung cancer, but may help distinguish benign nodules from metastases.

Introduction

Radiologists frequently encounter indeterminate lung nodules on CT examinations. Findings associated with an increased likelihood of malignancy, such as larger size, spiculated margins, and growth, support biopsy or resection. Nodules with less suspicious features, particularly smaller size, smooth rounded margins, and lack of growth typically are followed with serial CT examinations until features suspicious for malignancy develop (1, 2). Such surveillance of low risk lesions is intended to minimize the number of unnecessary invasive procedures for benign lesions, yet still allow diagnosis and treatment of the malignant lesions at an early stage.

Although CT surveillance may reduce the number of unnecessary invasive procedures, many still occur, with benign disease found in as many as 20% or more of pulmonary nodule resections (3–5). In addition, CT surveillance generates anxiety, effort, and increased medical costs even for benign lesions. For malignant nodules, CT surveillance delays biopsy, which could allow additional time for growth and metastasis. PET can be helpful in discriminating some benign and malignant lesions, but the metabolic activity in inflammatory and malignant nodules often overlaps, and PET is unreliable for nodules smaller than 1 cm diameter (6). Thus, despite an impressive array of technology, assessing the likelihood of malignancy quickly, accurately, and noninvasively remains a major challenge.

Previous investigators have developed reasonably successful prediction models to noninvasively discriminate benign from malignant nodules using combinations of clinical history and chest x-ray information (7–9). Other studies have found that computer-aided diagnosis using various quantitative features of pulmonary nodules on transverse CT sections can be moderately predictive of the benign or malignant nature of a nodule (10–12). However, a 2D transverse CT section samples only a small fraction of an entire pulmonary nodule, and 3D analysis of all spatial and attenuation information contained in the entire nodule might provide better classification. In this study, we hypothesized that three-dimensional (3D) volumetric CT quantitative analysis of lung nodules would improve the ability classify lung nodules compared to two-dimensional (2D) analysis. We also evaluated the additional impact of considering clinical information along with the CT data to enhance predictive ability.

Materials and Methods

Subject Eligibility

Patients ≥18 years of age scheduled for diagnostic and/or therapeutic resection of a pulmonary nodule ≤3 cm in greatest transverse dimension at our institution were identified from the daily clinic schedule of surgeons and enrolled by a study coordinator. Nodule size was confirmed by manual measurements on existing diagnostic CT images by one of the authors prior to enrollment (D.G.). The study was approved by the local institutional review board. Written informed consent was obtained from all subjects.

Clinical data collection

Subjects were interviewed using standardized forms to record history of hemoptysis, chronic obstructive pulmonary disease, emphysema, pulmonary fibrosis, other lung disease, and asbestos exposure. A detailed smoking history was obtained, including whether the patient had ever smoked; age began smoking; number of years smoked; amount smoked per day; and age of smoking cessation if applicable. Personal history and family history of cancer were elicited, including the cancer type and time since diagnosis.

CT imaging

All patients already had clinical chest CT scans performed prior to enrollment. For this study, an additional CT scan for quantitative nodule analysis was performed in each patient within one month prior to surgery. Scans were performed on a Definition 64 detector-row dual source scanner (Siemens), in single source mode, without intravenous contrast at full inspiration. To limit radiation exposure, the volumetric scan was targeted to the level of the nodule to be resected. In addition, thin-section images were obtained at 10 evenly spaced levels over the entire length of the lung to quantify emphysema (13).

The following technical parameters were used for both scanning protocols: 120 kVp; 130 mAs; automatic exposure control on; 0.33 second gantry rotation. For nodule scans, additional parameters included 0.6 mm detector collimation and pitch of 1.0. Nodule images were reconstructed at high spatial resolution, using a 102 mm field of view and 0.6 mm section thickness at 0.2 mm increments, to produce images with a voxel center spacing of 0.2 mm in each dimension. The B35f kernel was chosen for image reconstruction, after preliminary studies using a CTP657 phantom (The Phantom Laboratory) found that it provided the narrowest transition between materials of lung and soft tissue density among the available kernels that did not produce edge enhancement. Scans for quantifying emphysema were performed during a single, maximal inspiratory breath hold to acquire two contiguous 1 mm sections at each of the 10 evenly-spaced scan levels, at a field of view targeted to the widest diameter of the lungs.

Visual CT image analysis

The dedicated nodule CT scans were reviewed by a thoracic radiologist (D.G.), in conjunction with the preoperative clinical CT scans when necessary, to determine the lobe of each resected nodule and whether or not it was peripheral (within 2 cm of a pleural surface).

Quantitative CT image analysis

Nodule analysis

Nodules were segmented (separated) from the surrounding structures in the CT images with the software program 3D Slicer (version 4.0) (14) (Fig. 1). We used a 3D semiautomated iterative region growing method, starting from manually placed seed points, with region-growing according to whether the values of neighboring voxels were within a specified number of standard deviations of the region around the seeds; default values were used. Each segmented nodule was classified based on numerous quantitative CT parameters using software routines developed in MATLAB (Mathworks). Analyses of both 2D and 3D data were performed. For 2D analysis, only the single transverse section containing the greatest cross-sectional area of the nodule was analyzed. The 3D analysis used the entirety of the segmented nodule volume. The quantitative CT parameters extracted included multiple variables related to nodule size, shape, attenuation, texture, and margin (see Supplemental Data for full details):

• Size - Area and perimeter (2D); volume and surface area (3D).

• Shape - Circularity (2D); Sphericity (3D); long: short axis ratio; primary (2D and 3D), secondary (2D and 3D), and tertiary (3D) moments of inertia.

• Attenuation– Mean; median; standard deviation; variance; skewness; kurtosis; entropy.

• Texture – These parameters were selected to quantify local and regional differences in attenuation that describe nodule homogeneity, contrast, and complexity. Parameters included mean, standard deviation, variance, skewness, kurtosis, and entropy of the difference image; lacunarity of the segmentation image at box sizes of 1, 2, 4, 8, 16, 32, 64, and 128 voxels (2D and 3D); and coarseness, contrast, busyness, complexity, and texture strength, each at distances of both 1 and 2 voxels.

• Margin – These parameters were selected to quantify variations among nodules in surface contours. Parameters included the summed and mean distance to surface; normalized summed and mean distance to surface; fractal dimension of area or volume; fractal dimension of perimeter or surface; and gradient margin.

Figure 1.

Gray scale (top row) and binary segmented images of a 1.6 × 1.4 cm left lower lobe nodule. The diagnosis was adenocarcinoma.

Emphysema analysis

Emphysema was quantified by the proportion of the lungs having attenuation lower than −950 HU, using the software program Emphysema Profiler (VIDA Diagnostics) (15).

Pathologic diagnosis

Surgical pathology was obtained from resection specimens. Resected nodules were classified from pathology reports as benign, primary lung cancer, or metastasis.

Data analysis

Comparisons were made between primary lung cancers and benign nodules, primary lung cancers and metastases, and benign nodules and metastases. These comparisons were made individually, to see how each of these nodule categories differed from each of the others by quantitative CT classification, and because distinction between two of these categories is often the primary clinical concern involving indeterminate lung nodules. Clinical and quantitative CT variables were compared using t-tests for continuous variables and chi-square tests for categorical variables. Continuous variables were tested for normality using the Kolmogorov-Smirinov test, and those with non-normal distribution were log transformed. Variables for which two-tailed P values were less than 0.10 in the univariable analysis were examined for use in multivariable modeling, with the aim of testing at least one variable from each quantitative CT category (size, shape, attenuation, texture, and margins) in the models if possible, without using an excessive number of variables relative to the sample size. Fundamental parameters took precedence over those derived from the fundamental parameters (e.g., mean of the difference image took precedence over standard deviation of the difference image) if both had p values less than 0.10.

Multivariable models were created from clinical data alone, quantitative CT data alone, and combined clinical and quantitative CT data using stepwise logistic regression. A P value <0.05 was required for a variable to be retained in the model. Model performance was assessed by R² values and receiver operating characteristic curves using leave-one-out cross-validation. The areas under the receiver operating characteristic curves (AUC) were compared using the Mann-Whitney U test (16). Data were analyzed using SAS version 9.3.

RESULTS

Patients and Diagnoses

We enrolled 102 patients. Three patients did not undergo surgical resection after enrolling. Four patients were excluded due to technical inability of the segmentation program to separate the lung nodules from the chest wall. One patient was excluded due to pathologic and clinical uncertainty regarding whether the resected nodule was a primary lung cancer or a metastasis. Thus, 94 patients had complete data available for analysis. Two of these patients had two nodules in the same lung region that were scanned and resected, and each was treated as two independent samples. Thus, 96 nodules were analyzed.

The cohort consisted of 45 men and 49 women with a mean age of 62 yrs. ± 11 yrs. standard deviation (Table 1). Fifty-three of the nodules were primary lung cancer, 25 were metastases, and 18 were benign (Table 2). Eleven nodules were less than 1 cm in greatest transverse dimension, 44 were between 1 and 2 cm, and 41 were between 2 and 3 cm.

Table 1.

Comparisons of clinical parameters among patients with different nodule types.

			Lung Cancer(N=53)	Metastasis (N=25)	Benign (N=18)
Age, yearsa,b			65.0 +/− 10.1	60.1 +/− 11.1	57.7 +/− 10.9
Female, n(%)			31 (58.5%)	11 (44.0%)	8 (44.4%)
Caucasian, n(%)			48 (90.6%)	23 (92.0%)	17 (94.4%)
Ever smoked, n(%)a,b			48 (90.6%)	14 (56.0%)	9 (50.0%)
Among subjects ever smoked	Total years smokeda		36.4 +/− 14.0	26.9 +/− 15.0	26.9 +/− 15.0
	Pack years smoked		45.3 +/− 28.7	39.1 +/− 29.2	39.1 +/− 29.2
	Age started smoking, n(%)	>19	1 (11.1%)	1 (11.1%)	1 (11.1%)
		15–19	7 (77.8%)	7 (77.8%)	7 (77.8%)
		<15	1 (11.1%)	1 (11.1%)	1 (11.1%)
	Years quit smoking, n(%)	>14	4 (44.4%)	4 (44.4%)	4 (44.4%)
		1–14	2 (22.2%)	2 (22.2%)	2 (22.2%)
		<1	3 (33.3%)	3 (33.3%)	3 (33.3%)
Hemoptysis, n(%)			6 (11.3%)	0	2 (11.1%)
COPD, n(%)a,b			26 (49.1%)	2 (8.0%)	3 (16.7%)
Previous Cancer, n(%)c			14 (26.4%)	25 (100.0%)	6 (33.3%)
Family Cancer history, n(%)			40 (75.5%)	15 (60.0%)	12 (66.7%)
Nodule in upper/middle lobe, n(%)			37 (69.8%)	13 (52.0%)	11 (61.1%)
>2 cm from pleura, n(%)			7 (13.2%)	4 (16.0%)	1 (5.6%)
Pleural contact, n (%)			27 (50.9%)	12 (48.0%)	6 (33.3%)
Emphysema Index (%<−950 HU)			0.11 +/− 0.11	0.11 +/− 0.08	0.11 +/− 0.13
Upper Lung Index (%−950 HU)			0.12 +/− 0.13	0.10 +/− 0.10	0.11 +/− 0.15
Lower Lung Index (%−950 HU)			0.11 +/− 0.11	0.10 +/− 0.08	0.11 +/− 0.13

^aP<0.1, lung cancer vs. benign

^bP<0.1, lung cancer vs. metastasis

^cP<0.1, metastasis vs. benign

Table 2.

Nodule pathology.

Lung Cancer (N=53)	Metastases (N=25)	Benign (N=18)
Invasive adenocarcinoma (N=32)	Colon adenocarcinoma (N=5)	Granuloma (N=9)
Minimally invasive adenocarcinoma (N=6)	Head and neck squamous cell carcinoma (N=4)	Organizing pneumonia (N=4)
Squamous cell carcinoma (N=6)	1 each:	Hamartoma (N=3)
Carcinoid tumor (N=4)	Endometrial adenocarcinoma	Leiomyoma (N=1)
Small cell carcinoma (N=3)	Ethmoid spindle cell carcinoma	Dystrophic calcification (N=1)
Undifferentiated or unspecified (N=2)	Forearm nerve sheath sarcoma
	Knee synovial cell sarcoma
	Lower extremity fibrosarcoma
	Lung squamous cell carcinoma
	Ovarian adenocarcinoma
	Pancreatic neuroendocrine carcinoma
	Parotid melanoma
	Penis leiomyosarcoma
	Pleural fibrosarcoma
	Thigh hemangiopericytoma
	Thigh malignant fibrous histiocytoma
	Thigh melanoma
	Thyroid poorly differentiated carcinoma
	Uterine leiomyosarcoma.

Primary Lung Cancer vs. Benign

Clinical Model

There were univariable differences at the P<0.10 level between groups (Table 1) for age, whether the patient had ever smoked, total years smoked, pack years smoked, and history of COPD. In stepwise logistic regression using these clinical parameters, the only significant parameter for distinguishing lung cancer from benign nodules was whether the patient had ever smoked (Model R²=0.24; AUC=0.70) (Table 3).

Table 3.

Multiple logistic regression models for distinguishing primary lung cancer from benign nodules.

Model	Variable	ORa	95% CI	P value	R2	AUCb
Clinical	Ever smoked	9.6	2.6–35.4	0.001	0.24	0.70
2D CT	Areac	3.39	1.18–9.77	0.024	0.30	0.79
	Circularity (unit=0.01)	0.94	0.89–1.00	0.049
3DCT	Difference image meanc	0.28	0.11–0.71	0.008	0.32	0.80
	Sphericity (unit=0.01)	0.91	0.85–0.98	0.010
Clinical-2D CT combined	Ever smoked	14.0	2.9–68	0.001	0.43	0.86
	Area	6.1	1.8–20.6	0.021
Clinical-3D CT combined	Ever smoked	11.5	2.2–59.9	0.004	0.47	0.88
	Difference image mean	0.22	0.07–0.69	0.01
	Sphericity	0.91	0.85–0.99	0.021

^aOdds ratio (OR) >1 indicates higher likelihood of lung cancer as CT parameters increase one unit. Change unit for circularity and sphericity is 0.01.

^bArea under receiver operating characteristic curve.

^cLog transformed values

2D CT Model

Of the 2D nodule parameters that differed at the P<0.10 level (Supplement Table S1), 10 were used in multivariable modeling: area, circularity, attenuation entropy, difference image mean, difference image entropy, lacunarity at box size=32 voxels, coarseness at distance=1 voxel, distance to surface normalized mean, fractal dimension of perimeter, and gradient margin. Area and circularity remained significant in the stepwise logistic regression model (Model R²=0.30; AUC=0.79) (Table 3).

3D CT Model

Of the 3D nodule parameters that differed at the P<0.10 level (Supplement Table S2), those used in multivariable modeling were the same as or analogous to those used in 2D modeling: volume, sphericity, attenuation entropy, difference image mean, difference image entropy, lacunarity at box size=32 voxels, coarseness at distance=1 voxel, distance to surface normalized mean, fractal dimension of surface, and gradient margin. Sphericity and difference image mean remained significant predictors of lung cancer in the stepwise logistic regression model (model R²=0.32; AUC=0.80) (Table 3). The AUC of the 3D CT model was not statistically different from the AUC of the 2D CT model (p=0.78).

Combined Clinical-CT Models

In the combined model using parameters of the stepwise clinical and 2D CT models, whether the patient had ever smoked and nodule area were both significant predictors of lung cancer (model R²=0.43; AUC=0.86) (Table 3). The AUC for the combined clinical-2D CT model was not statistically different from the AUC for the 2D CT model alone (P=0.24). In the combined model using parameters of the stepwise clinical and 3D CT models, whether the patient had ever smoked, sphericity, and difference image mean were all significant predictors (model R²=0.47; AUC=0.88) (Table 3). The AUC of the combined clinical-3D CT model was not statistically different from the AUC of the 3D CT model alone (P=0.12). There was no significant difference between the AUC values of the combined clinical-2D CT model compared to the combined clinical-3D CT model (P=0.71).

Primary Lung Cancer vs. Metastasis

Clinical Model

There were univariable differences at the P<0.10 level (Table 1) for age, whether the patient had ever smoked, total years smoked, pack years smoked, history of COPD, and history of previous cancer. Because there were no patients with metastatic disease who did not have a history of previous cancer, we were unable to assess cancer history in the stepwise logistic regression model. Whether the patient had ever smoked and a history of COPD were significant variables in the stepwise model (R²=0.31, AUC=0.77) (Table 4).

Table 4.

Multiple logistic regression models for distinguishing primary lung cancer from metastases.

Model	Variable	ORa	95% CI	P value	R2	AUCb
Clinical	COPD	7.09	1.43–35.2	0.017	0.31	0.77
	Ever smoked	4.03	1.13–14.4	0.031
2D CT	Circularity (unit=0.01)	0.94	0.89–0.98	0.006	0.34	0.81
	Difference image entropy c	0.49	0.02–1.00	0.047
3DCT	Lacunarity at box size=32 (unit=0.01)	4.1	1.8–9.3	0.001	0.40	0.83
Clinical-2D CT combined	COPD	10.1	2.04–50.2	0.001	0.43	0.85
	Circularity (unit=0.01)	0.92	0.88–0.97	0.005
Clinical-3D CT combined	Lacunarity at box size=32 (unit=0.01)	1.10	1.04–1.16	0.001	0.56	0.90
	COPD	15.5	2.65–90.5	0.002

^aOdds Ratio (OR) >1 indicates higher likelihood of primary lung cancer as CT parameters increase one unit. Change unit for circularity and lacunarity is 0.01.

^bArea under receiver operating characteristic curve

^cLog transformed values

2D CT Model

Of the nodule parameters that differed between groups at the P<0.10 level (Supplement Table S3), 10 were used in logistic regression modeling: area, circularity, attenuation entropy, difference image mean, difference image entropy, lacunarity at box size=32 voxels, complexity at distance=2 voxels, distance to surface normalized mean, fractal dimension of perimeter, and gradient margin. Circularity and difference image entropy were significant in the final model (R²=0.34, AUC=0.81) (Table 4).

3D CT Model

Of the 3D nodule parameters that differed at the P<0.10 level (Supplement Table S4), 10 of those used in multivariable modeling were the same as or analogous to those used in 2D modeling: volume, sphericity, attenuation entropy, difference image mean, difference image entropy, lacunarity at box size=32 voxels, complexity at distance=1 voxel, distance to surface normalized mean, fractal dimension of surface, and gradient margin. Coarseness at distance=1 voxel also was included. In the final model, only lacunarity at box size=32 voxels remained significant at P<0.05 (R²=0.40, AUC=0.83) (Table 4). The AUC for the 3D CT model was not statistically different from the AUC of the 2D CT model (P=0.29).

Combined CT-Clinical Models

A history of COPD and circularity were significant in the combined clinical-2D CT model (model R²=0.43; AUC=0.85). The AUC of the combined clinical-2D CT model was not statistically different from the AUC of the 2D CT model alone (P=0.16). A history of COPD and lacunarity at box size=32 remained significant in the combined clinical-3D CT model (model R²=0.56; AUC=0.90) (Table 4). The AUC of the combined clinical-3D CT model was significantly higher than the AUC of the 3D CT model alone (P=0.04). The AUC values of the combined clinical-2D CT and combined clinical-3D CT models were not significantly different (P=0.13).

Benign Nodule vs. Metastasis

Clinical Model

The only clinical parameter that differed at P<0.10 between those with benign nodules and those with metastases was a history of previous cancer. However, because there were no patients with metastatic disease who did not have a history of previous cancer, we were unable to assess cancer history by logistic regression modeling.

2D CT Model

Three of the nodule parameters that differed between groups at the P<0.10 level (Supplement Table S5) were used in 2D multivariable modeling: attenuation skewness, busyness at distance=2 voxels, and texture strength at distance=2 voxels. None of these parameters were significant at the P<0.05 level in stepwise logistic regression modeling; attenuation skewness was the strongest 2D parameter (p=0.062; model R²=0.12; AUC=0.68).

3D CT Model

Nodule parameters that differed in 3D univariable comparisons at the P<0.10 level (Supplement Table S6) and were used in 3D multivariable modeling included attenuation skewness, standard deviation of difference image, skewness of difference image, lacunarity at box size=64, contrast at distance=2 voxels, busyness at distance=2 voxels, texture strength at distance=2 voxels, and complexity at distance=2 voxels. Skewness of the difference image (P=0.01) and lacunarity at box size=64 (P=0.03) remained significant predictors in the stepwise logistic regression model (model R²=0.43; AUC=0.84).

Combined Clinical-CT Models

Due to the lack of testable clinical parameters, combined clinical-CT models for distinguishing benign nodules and metastases could not be created.

Discussion

In this initial study of a relatively small cohort, there was no difference in the ability of 2D and 3D analysis approaches to distinguish primary lung cancer from benign nodules or from metastases, and 3D analysis was superior to 2D analysis for distinguishing benign nodules and metastases. Thus, whether 3D nodule analysis is superior may depend on the types of nodules one wishes to discriminate, at least with the quantitative CT parameters tested. Whether 3D volumetry is superior to the current standard of 2D linear measurements for assessing change in size over time was not evaluated in this study.

Although it is disappointing and somewhat surprising that we did not find improvement in the ability to distinguish primary lung cancer from other nodule types with 3D analysis, 2D analysis may have some positive practical implications if equivalent to 3D. For example, quantitative 2D CT nodule classification could be applied to data sets that are not amenable to 3D analysis, such as routine, non-thin section clinical scans on which indeterminate nodules are incidentally detected. It also could be applied in research using cases in existing publicly-available lung CT research databases that do not have many thin section CT cases needed for 3D analysis, such as those of the National Lung Screening Trial (17) and the Lung Imaging Database Consortium (18), without undue concern that diagnostic performance is being compromised. Reduced segmentation and nodule analysis time may be a minor practical benefit of the 2D approach. In addition, equivalence of 2D and 3D model performance would increase confidence that the predictive potential of non-contrast quantitative CT for lung nodule classification has been accurately estimated by 2D models developed previously.

Our quantitative CT models were moderately successful in discriminating nodule types, with AUC values ranging from 0.79–0.84. To illustrate the performance in other terms, in the 3D model distinguishing lung cancer from benign nodules (AUC=0.80), a probability cutoff for malignancy of 0.58 was associated with 87% sensitivity, 50% specificity, and correct predictions in 78% of cases. Other studies have found AUC values from 2D CT data prediction models of 0.83 to 0.92 (12, 19). A model that combined 2D and 3D nodule features achieved an AUC of 0.94 (20), while an entirely 3D classification system found an AUC of 0.83 for distinguishing 44 malignant from 52 benign nodules (21). In these comparison studies, pathologic proof was not obtained for most benign nodules and some presumed malignant nodules, whereas proof was obtained for all cases in our study. The nodules that were benign in our study may have been more challenging to classify than in other studies since they were of high enough concern to resect.

Analogous 2D and 3D input variables were used to create the multivariable models for distinguishing primary lung cancer from benign nodules and from metastases. However, the resultant 2D and 3D models did not contain only analogous variables, suggesting that the relative influence of these variables may differ depending on whether nodules are analyzed in 2D or 3D, or that some of the variables may have been correlated with each other. In building the models for distinguishing metastases from benign nodules, the univariable analyses did not provide analogous input variables for multivariable regression, and only the 3D model was statistically significant. Additional notable features are that all of the models contain parameters that describe nodule texture, which may be difficult to evaluate visually, and that the texture feature of lacunarity was a discriminator of metastases from both primary lung cancer and benign nodules.

Previous clinical models developed using demographic, smoking, and medical history information also included some subjectively assessed nodule features, and found AUC values in the range of 0.79–0.89 (7, 9, 22). The clinical model in our study for differentiating lung cancer and benign nodules had a lower AUC of 0.70, but did not include any nodule features. The AUC values of our combined clinical-CT models were greater than those for our clinical or CT models alone; most of the differences were not statistically significant, but given the consistently greater values for the combined models there may be a trend that our sample size was not large enough to confirm with statistical significance.

Because there were no patients with lung metastases who did not have a history of extrathoracic malignancy, the predictive value of a history of previous malignancy could not be evaluated in the multivariable clinical models. The expected frequency of patients who meet this condition is low, so a much larger sample would be needed to quantify the discriminatory power of a history of previous malignancy. Because of an expected strong association between a history of malignancy and a metastatic etiology of a lung nodule, the performance of the clinical and combined models for distinguishing metastases from primary lung cancer and benign nodules was probably underestimated.

There are several limitations of this study. The relatively high ratio of CT variables tested to nodules analyzed, and potential intercorrelation among some CT variables, could have led to the identification of some statistically significant variables by chance. This is a common problem in lung nodule classification efforts, and multiple approaches to reducing the number of variables have been proposed (23). For this reason, we limited the number of variables used in the multivariable modeling. This issue has limited bearing on the results of our study, as it was not designed to develop and validate definitive prediction models, but rather to compare the performance of 2D and 3D models generated in the same manner using analogous variables. While the proportion of nodule types in the sample might influence performance of the models, the effect of this on the comparison of 2D and 3D approaches likewise should be limited. The size of our cohort was relatively small, reflecting the volume and proportion of nodule types found in a single academic thoracic surgery department over two years, so the results should be verified in other populations.

We conclude that 3D quantitative nodule analysis may not improve the ability to distinguish primary lung cancer from benign nodules or metastases compared to 2D analysis, but may improve the discrimination of benign nodules from metastases. To improve model performance, exploring other strategies may be worthwhile, such as devising quantitative CT descriptors with more discriminatory power, using other statistical or machine learning methods for feature selection and modeling, or incorporating growth rate into a model. Whether a validated nodule classification model that provides a probability of malignancy would improve patient management decisions will need to be assessed in a clinical trial setting. However, development of a widely generalizable, robust model will likely require a database of nodules much larger than those that have been previously used. Since evaluation of virtually all lung nodules includes a CT examination, there is realistic potential for application in clinical practice. Opportunity and need for this may increase now that lung cancer screening for high-risk individuals is recommended by the United States Preventive Services Task Force (24).