Radiology
Binsheng Zhao, DSc Lawrence H. Schwartz, MD Chaya S. Moskowitz, PhD Liang Wang, MD Michelle S. Ginsberg, MD Cathleen A. Cooper, BS Li Jiang, MS John P. Kalaigian, BS Published online before print 10.1148/radiol.2343040020 Radiology 2005; 234:934 –939
Pulmonary Metastases: Effect of CT Section Thickness on Measurement—Initial Experience1 PURPOSE: To assess the effect of commonly used computed tomographic (CT) section thicknesses on metastatic tumor measurements calculated with unidimensional, bidimensional, area, and volumetric methods.
1
From the Departments of Radiology (B.Z., L.H.S., L.W., M.S.G., C.A.C., L.J., J.P.K.) and Epidemiology and Biostatistics (C.S.M.), Memorial Sloan-Kettering Cancer Center, 1275 York Ave, New York, NY 10021. Received January 7, 2004; revision requested March 9; revision received May 3; accepted June 2. Supported in part by grants from William H. Goodwin and Alice Goodwin and the Commonwealth Foundation for Cancer Research and The Experimental Therapeutics Center of Memorial Sloan-Kettering Cancer Center. Address correspondence to B.Z. (e-mail:
[email protected]).
Authors stated no financial relationship to disclose.
MATERIALS AND METHODS: Analysis and data collection were approved by the Institutional Review Board, with waived informed patient consent. Forty-two pulmonary metastases in 10 patients (three men and seven women; age range, 43– 83 years; mean age, 65.4 years) were analyzed on CT scans obtained with 3.75-, 5.0-, and 7.5-mm section thicknesses. The lesions were automatically delineated by using a three-dimensional multicriteria segmentation algorithm. Unidimensional (the largest diameter), bidimensional (the product of the two maximal perpendicular diameters), maximal cross-sectional area, and volumetric measurements were automatically obtained for each pulmonary lesion on each section thickness. Means and variances were calculated, and the differences across the three section thicknesses for each of the four measurements were studied by using linear mixed-effects models. The Levene test was used to study the equality of variances. RESULTS: Differences in the means for unidimensional, bidimensional, and area measurements were significant between a section thickness of 3.75 and 5.0 mm (unidimensional, P ⫽ .05; bidimensional, P ⫽ .05; area, P ⫽ .01) and 3.75 and 7.5 mm (unidimensional, P ⫽ .06; bidimensional, P ⫽ .03; area, P ⫽ .02), but not 5.0 and 7.5 mm. There was a significant difference in volumetric measurement as section thickness decreased from 7.5 to 5.0 mm (P ⬍.001) and from 7.5 to 3.75 mm (P ⬍ .001). Although there was a slight trend for differences in the variances across section thickness for each measurement, none of the differences were significant. CONCLUSION: Volumetric tumor measurements change with a reduction in section thickness from 7.5 to 5.0 and 3.75 mm. For unidimensional measurement, no change was found when thickness decreased from 7.5 to 5.0 mm. ©
Author contributions: Guarantors of integrity of entire study, L.H.S., B.Z.; study concepts, L.H.S., B.Z., M.S.G.; study design, L.H.S., B.Z., C.S.M.; literature research, B.Z., L.H.S.; clinical studies, L.W., L.J., B.Z., L.H.S., M.S.G.; data acquisition, J.P.K., C.A.C., L.J.; data analysis/interpretation, C.S.M., L.H.S., B.Z.; statistical analysis, C.S.M., L.H.S., B.Z.; manuscript preparation, editing, and final version approval, B.Z., L.H.S., C.S.M.; manuscript definition of intellectual content and revision/review, B.Z., L.H.S. ©
RSNA, 2005
934
RSNA, 2005
Change in tumor size observed with serial radiologic examinations has been the traditional standard for the assessment of therapeutic response in cancer patients (1–7). Accurate and reproducible measurements of tumors are thus of vital importance, as the results can influence a patient’s ongoing treatment plan or may determine the fate of an investigational drug under evaluation. One of the following four image-based tumor measurements may be performed: unidimensional, bidimensional, area, or volumetric. With the unidimensional measurement, the largest diameter of a tumor on a section with the largest cross-section of the tumor is measured. With the bidimensional measurement, the cross product of the largest tumor diameter and the maximal diameter perpendicular to it are measured. The area is calculated by measuring the maximal cross-sectional area of the tumor on a single computed tomographic (CT) section, while the volume is calculated by summing the tumor areas across all the sections that contain the tumor. The evolution of medical imaging technology is toward acquisition of images with faster speed and at higher resolution, with the goals of reducing image artifacts, improving
Radiology
detection of small lesions, and assessing subtle changes of disease response or progression. Furthermore, the digital environment of hospitals, including picture archiving and communication system, allows not only superior manual measurements on workstations but also the integration of computerized measurements into the system. Thin-section multi– detector row CT and volumetric magnetic resonance (MR) imagers generate large quantities of data, and benefits may be realized in the detection of smaller lesions and in the ability to reconstruct images in multiple planes. A practical issue has been raised as to whether thin-section images are always needed. In other words, what would be the appropriate image resolution for a certain application, such as therapy response assessment, where rules governing response and progression based on change in lesion size have already been established? In the current clinical practice and in clinical trials, the assessment of the effectiveness of a drug relies on measuring tumor growth or regression by using unidimensional or bidimensional measurements. The historical use of unidimensional and bidimensional measurements is based on the fact that these measurements can be easily performed by a radiologist with handheld calipers for hardcopy images or with electronic calipers at display monitors. Rules governing the regression or growth of a tumor with unidimensional and bidimensional measurements have been established and are routinely used, and suggestions for CT section thickness relative to lesion size have been proposed but not practically tested (1,2). In addition, some groups have advocated the use of volumetric assessment of a tumor (8). While there is debate regarding the appropriate methods of assessing tumor response, little attention has been paid to the methods of data acquisition. In general, routine site-defined CT scanning protocols are used for phase II and III oncology trials. Section thickness considerations are limited to narrow ranges of possible acquisitions, since frequent serial scanning is generally performed on large (chest, abdomen, and pelvis) areas. Thus, the purpose of our study was to assess the effect of commonly used CT section thicknesses on metastatic tumor measurements calculated with the following four methods: unidimensional, bidimensional, area, and volumetric. Volume 234
䡠
Number 3
MATERIALS AND METHODS Patient Image Selection Retrospective data collection and analysis were approved by the Institutional Review Board, with waived informed patient consent. Clinically acquired multi– detector row (eight-row) CT images (LightSpeed QX/i; GE Medical Systems, Milwaukee, Wis) of the chest were retrieved from the hospital picture archiving and communication system and were transferred to our research picture archiving and communication system archive via the hospital network system. The Digital Imaging and Communications in Medicine images were stored in the archive, with patient identification information (name, medical record identification, etc) removed. These images were then transferred to a workstation (Sun Ultra 10; Sun Microsystems, Palo Alto, Calif) for further processing. CT scans obtained from 10 patients (three men, seven women; average age, 65.4 years; range, 43– 83 years) with documented pulmonary metastases (at biopsy or interval growth at sequential scanning) were randomly selected from daily clinical practice. The primary tumors were lung (n ⫽ 2), melanoma (n ⫽ 1), sarcoma (n ⫽ 1), renal (n ⫽ 2), colorectal (n ⫽ 3), and breast (n ⫽ 1). The acquisition protocol for this assessment was a beam pitch of 0.75, 120 kVp, 190 – 290 mAs, pitch of 1.35, and 7.5-mm collimation with 7.5-mm reconstruction interval. In addition to the 7.5-mm reconstruction interval, raw data were reconstructed with additional two narrower intervals, 5.0 and 3.75 mm. Nodules larger than 5 mm were chosen to simulate the actual requirements of the Response Evaluation Criteria in Solid Tumors and of the World Health Organization, as well as the routine daily practice of oncologic radiologists. Forty-two such pulmonary metastases were identified in the 10 patients and were included in this evaluation.
Automated Segmentation and Quantification An automated three-dimensional multicriteria algorithm was developed previously, making use of the differences in opacity and shape to separate pulmonary lesions from their surrounding tissues on CT images (9 –11). By using this algorithm, a rectangular region of interest, which tightly enclosed the lesion under
study, was outlined manually by an operator (L.W. or B.Z.) on one CT section and then propagated automatically to all other sections. To reduce the partial volume effect of CT scanners and to simplify definitions of some three-dimensional operations, the lesion volume of interest was supersampled by using a trilinear interpolation to an isotropic scale in the x-, y-, and z-axes. An initial high attenuation value was determined on the basis of the opacity distribution of the voxels in the lesion volume of interest. The threshold was decreased stepwise until the one subsequent threshold resulted in substantial changes to the gradient strength for the lesion surface (average opacity changes on the lesion surface). If the lesion identified at this threshold level possessed a compact shape, the segmentation terminated. Otherwise, a series of morphologic operations would be applied, starting with a small filter size. As soon as the lesion showed a compact shape, the erosion ceased. Use of the shape criterion introduced to the algorithm could efficiently detach the surrounding high-opacity structures such as blood vessels from pulmonary lesions. The algorithm had been validated with small lung nodule phantoms and kidney stone phantoms on high-spatial-resolution helical CT images and achieved accurate volumetric measurements (12,13). Once a lesion was separated from its surrounding tissues, its maximal diameter, two perpendicular maximal diameters, maximal area, and volume could be automatically calculated. Explicitly, the area of a lesion on a CT section was the number of lesion pixels on that section multiplied by two in-plane resolutions in x- and y-axes. The area measurement was defined as the largest in-plane area among the lesion areas on all sections containing the lesion. Lesion volume was the sum of lesion areas on each of the lesion sections multiplied by the scan transverse resolution. To obtain uniand bidimensional diameters of a lesion, the contour of the cross-sectional lesion having the maximal area was extracted. The longest distance of any two pixels on the contour was automatically taken as the largest diameter of the lesion. Along the line of the maximal diameter, perpendicular diameters were then determined, and the longest diameter multiplied by the in-plane resolution (in x- or y-axis) was defined as the largest perpendicular diameter. The Figure is an example of the automated segmentation and quantification of a lung nodule at multi– detector row
Pulmonary Metastases: Effect of CT Section Thickness
䡠
935
Radiology Automated segmentation and quantification of lung nodules on transverse CT images. (a) One CT section with a manual selection of a nodule region of interest. Area surrounded by white line indicates the region of interest. (b) Consecutive CT sections (1– 8) of the nodule regions of interest. Nodule contours obtained with the multicriteria segmentation algorithm were overlapped on the images. The nodule on the fourth section was found to have the maximal area, and two perpendicular maximal diameters were determined thereafter. (c) Intermediate segmentation result by using the gradient criterion. (d) Final segmentation result using both the gradient and shape criteria.
CT. A 31.9 ⫻ 24.2-mm nodule was surrounded by blood vessels in the right lung (within the white outline in part a). Part c displays an intermediate segmentation result, which was obtained by using only the gradient strength criterion, where some of the contiguous vessels were falsely segmented as part of the nodule. After the shape criterion was applied, the vessels were successfully removed from the nodule, whereas the details of the nodule surface were retained (Figure, part d). Part b shows the automatically obtained nodule contours overlapped on the nodule images, the maximal area section, and the two perpendicular diameters. In 10 patients, 42 nodules, each being reconstructed with three different section thicknesses, were evaluated by using the four measurement techniques. For the visual inspection of segmentation results, the software allowed us to view each of the sections containing the segmented nodule contour overlapped onto the original nodule images in a cine model, as well as the two largest diameters that were placed within the segmented nodule. Furthermore, the nodules could be viewed three dimensionally (Figure, part d). The segmentation results were visually reviewed (L.W., B.Z.) for the correctness of the nodule contour and the two diameters. 936
䡠
Radiology
䡠
March 2005
Statistical Analysis
RESULTS
To determine whether section thickness affects the different measurements (unidimensional, bidimensional, area, and volumetric), the means and variances of the different section thicknesses were compared for each of the four measurements. Linear mixed-effects models (14) were used to compare the differences in the means of the three section thicknesses, while accounting for the correlated nature of the data due to multiple nodules and multiple scans for each patient. These comparisons were made by using the natural logarithm scale to satisfy the assumptions of linear mixed-effects model methods. Multiple comparisons were adjusted by using the method of Tukey (15). To compare variances, permutation tests with a modification of the Levene test for equality of variances proposed by Brown and Forsythe were used (16,17), where we carefully permuted the data to reflect the correlation. For all analyses, we considered P ⱕ .05 to indicate a significant difference. The linear mixed-effects model analysis was performed with SAS software (version 8.00; SAS Institute, Cary, NC), and variance comparisons were performed with S-Plus 2000 software (MathSoft, Seattle, Wash).
A summary of the measurements across lesions is given in Table 1. The table indicates that different section thicknesses may in fact yield different tumor measurements, but the differences do not appear to be very large, especially for the unidimensional, bidimensional, and area measurements. Comparison of the differences in the means of the log-transformed data between the three section thicknesses and the unidimensional, bidimensional, area, and volumetric measurements shows that the overall test of whether section thickness is associated with each measurement is always significant (P ⬍ .001, Table 2). This result indicates that significantly different measurements may be obtained when different section thicknesses are used. To more thoroughly explore where the differences occurred, the pair-wise comparisons between the three section thicknesses within each measure showed a trend for the unidimensional, bidimensional, and area measurements (Table 2). While in some cases the differences were only marginally significant, the data do suggest that there are differences between a section thickness of 3.75 and 5.0 mm (unidimensional, P ⫽ .05; bidimenZhao et al
Radiology
TABLE 1 Summary of Data across Lesions with Use of Three Section Thicknesses and Four Measurements Section Thickness (mm)
Mean
Range
SD
Mean
Range
SD
Mean
Range
SD
Mean
Range
SD
3.75 5.0 7.5
14.1 13.8 13.8
(7.3–31.9) (6.9–30.8) (6.3–32.0)
4.6 4.5 4.8
172.1 165.8 166.6
(36.6–771.2) (29.0–757.5) (28.5–792.0)
126.1 121.3 130.5
124.8 120.4 121.2
(27.1–545.4) (21.1–548.6) (20.7–570.2)
89.2 88.3 92.2
1259.8 1287.9 1547.9
(117.5–8181.4) (105.6–8507.9) (132.4–10993.6)
1349.4 1361.2 1779.5
Bidimensional (mm2)†
Unidimensional (mm)*
Area (mm2)
Volumetric (mm3)
Note.—SD ⫽ standard deviation. * Measurement of maximal diameter. † Measurement of cross product.
TABLE 2 Differences in the Means of the Log-transformed Data between Section Thicknesses and the Four Measurements Bidimensional†
Unidimensional* Section Thickness (mm) 3.75 vs 5.0 3.75 vs 7.5 5.0 vs 7.5
Area
Volumetric
Difference in Means
P Value‡
Difference in Means
P Value‡
Difference in Means
P Value‡
Difference in Means
P Value‡
0.025 0.030 0.006
.052 .058 .850
0.042 0.057 0.015
.051 .025 .657
0.043 0.045 0.002
.008 .018 .986
0.044 0.220 0.176
.267 ⬍.001 ⬍.001
Note.—Overall F test performed to determine the effect of section thickness on the various measurements was ⬍0.001 for all measurements. * Measurement of maximal diameter. † Measurement of cross product. ‡ P value adjusted for multiple pair-wise comparisons.
sional, P ⫽ .05; area, P ⫽ .01) and between 3.75 and 7.5 mm (unidimensional, P ⫽ .06; bidimensional, P ⫽ .03; area, P ⫽ .02) but that there is no significant difference between a section thickness of 5.0 and 7.5 mm. In other words, for these three particular measurements, there does not appear to be a benefit of increasing the resolution from 7.5 to 5.0 mm, but use of a section thickness of 3.75 mm could result in substantially different measurements of tumor size. This effect is less pronounced with the unidimensional measurement than with bidimensional and area measurements. In contrast, the results of the volumetric measurement do not follow the pattern of the other three measurements. Section thicknesses of 3.75 and 5.0 mm do not produce significantly different volumetric measurements, but a section thickness of 7.5 mm does produce significantly different measurements than do the other two section thicknesses (P ⬍ .001 for both comparisons, Table 2). In an attempt to determine whether resolution affects the precision of the measurements, comparison of the variances of the three section thicknesses with each measurement was made (Table 3). For the unidimensional, bidimensional, and area measurements, the variance in tumor measurements is similar across the section thickness. For the voluVolume 234
䡠
Number 3
TABLE 3 Variances of the Section Thicknesses for Each of the Four Measurements Section Thickness (mm)
Unidimensional*
Bidimensional†
Area‡
Volumetric§
3.75 5.0 7.5
21.0 20.0 23.5
15 888 14 701 17 037
7959 7791 8504
1 820 919 1 852 858 3 166 758
* Measurement of maximal diameter, P ⫽ .978. † Measurement of cross product, P ⫽ .980. ‡ P ⫽ .988. § P ⫽ .593.
metric measurement, the data suggest that a 7.5-mm section thickness may result in higher variability of the measurements, but the difference in variation is not significant (P ⫽ .60).
DISCUSSION Accurate and reproducible assessment of the therapeutic response to chemotherapy and radiation therapy is critical in judging the success or failure of a therapy for a given patient or in a clinical trial setting. There are a number of different methods of assessing therapy response. The World Health Organization has recommended a standardized approach to assess response, wherein the largest diameter and the largest perpendicular di-
ameter of each selected tumor deposit are measured on CT or MR images. Recently, the Response Evaluation Criteria in Solid Tumors group has issued new guidelines for evaluating tumor response, in which only the largest diameter of a tumor (unidimensional measurement) is used (6). Response evaluation in general is divided into the following categories: complete response, disappearance of all tumors; partial response, a decrease of 30%–50% in the sum of lesions; and disease progression, an increase of 20%–25% in tumor size, depending on which method is used. Study findings have shown substantial variability in therapeutic response assessment. Both intraobserver and interobserver variations in tumor measurement
Pulmonary Metastases: Effect of CT Section Thickness
䡠
937
Radiology
exist for a variety of reasons, such as size, irregular shape, and poorly defined margins of a lesion; the phase of intravenous contrast agent administration and the exact levels through which a lesion is scanned; and the measurement technique used (eg, handheld calipers, electronic calipers, or automated techniques) (18 –23). Frequently in clinical trials, patients may be scanned with a variety of scanners and a variety of image acquisition protocols. For example, it is not uncommon for a patient’s baseline study to be performed at one institution and, when the patient enrolls in a protocol, to have the follow-up scans obtained at another institution, which may use different equipment and imaging protocols. The effect of this change needs to be assessed and should be evaluated with reference to the type of measurement obtained. Our results suggest that, especially with unidimensional and bidimensional tumor measurements, there is no difference in varying section thicknesses between 7.5 and 5.0 mm. It is not surprising that when volumetric measurements are used, the differences in volume between various section thickness are great. A recently published article (24) reported that volumetric overestimation of simulated lung tumors on spiral CT images ranged from 15% (38.1 mm in diameter, 2-mm section thickness) to 278% (12.7 mm in diameter, 10-mm section thickness) when a perimeter method (ie, a radiologist manually drew the nodule contour on each section containing the nodule, and the area enclosed by the perimeter was calculated by using computer software) was used. For the first time, to our knowledge, the authors, using the phantom data, derived error-reduction equations to compensate for the overestimation caused by the different section thicknesses and varying nodule sizes. Findings of their preliminary study showed promise of such compensatory equations in reducing the overestimation with clinical subjects (24). Other factors may have a great effect on lesion size measurements. For instance, the evaluation of focal hepatic metastases will vary greatly depending on the use of a contrast agent and the phase of contrast agent administration. Other soft-tissue mass measurements are affected by lesion border delineation. Lesions with relatively little difference in contrast between the lesion and the surrounding parenchyma may be different to define and measure. It is for this reason that we choose to assess the effect of sec938
䡠
Radiology
䡠
March 2005
tion thickness on pulmonary nodules, which have relatively well-defined borders. In our experimental study, we attempted to reduce the measurement errors caused by the partial volume effect of a CT scanner and by the systematic sampling by using a super-resampling technique. Nevertheless, theoretic analysis of these errors is beyond the scope of this study and can be found in other published work (25,26). Fully automatic computerized methods, including the unidimensional, bidimensional, area, and volumetric measurements, have been implemented and applied to the size estimation of lung nodules on CT images reconstructed at different section thicknesses. One of the advantages of the segmentation algorithm used in this work is its automatic, objective, and consistent way to determine the optimal thresholds to separate nodules from the surrounding tissues. Although the algorithm requires a manual initiation by placing a region of interest that tightly encloses the nodule to be segmented, it can maximally minimize, if not totally eliminate, variations in the segmentation results caused by different selections of the nodule region of interest. To maximally reduce false segmentation results caused by the surrounding tissues of similar opacities, we tried to (a) choose isolated nodules whenever possible, that is, nodules that were not attached to any surrounding tissues, or (b) maximally exclude the contiguous surrounding tissues of similar opacities, such as blood vessels, when drawing the rectangular region of interest. Because of the nature of the algorithm—that is, its use of the average gradient calculated from the nodule surface to determine an optimal threshold—it favors sharper CT images (eg, images reconstructed with the lung convolution kernel), particularly when nodules are small. However, not much attention was paid to the choice of CT reconstruction algorithm since (a) the same imaging protocol was used throughout the study and (b) the nodule sizes of interest were relatively large. How the CT reconstruction algorithm affects the segmentation algorithm is beyond the scope of this study. Because the in-plane CT resolution is relatively high (0.6 – 0.8 mm) compared with the size of the nodule of interest (⬎5 mm), inaccuracy in the measurement is primarily caused by the lower resolution along the z-axis. Because of partial volume averaging, the largest diameter, perpendicular largest diameters, area, and
volume of a nodule may have different attenuation values when measured on an image with a thinner section thickness than when measured on an image with a thicker section thickness. A major limitation of our study was the lack of a standard for the size of these in vivo nodules. However, it was not necessarily the purpose of our study to prove the accuracy of the techniques but to evaluate the differences in the results obtained in terms of both absolute attenuation values and variances. Our data indicate that the results obtained depend on the type of measurement being performed. We found that measurements performed in a single plane are less “sensitive” to section thickness; thus, there is little advantage in decreasing section thickness from 7.5 to 5.0 mm. Volumetric measurements, however, demonstrate a greater change between 7.5- and 5.0-mm and 7.5- and 3.75-mm section thickness. These data should affect the design and implementation of the imaging portion of clinical trials. In this article, we have reported results of our initial experience. Further analysis with a large sample size and different types of lesions and imaging modalities is required to confirm our results and generalize these guidelines. These data must be interpreted carefully with regard to the actual nodule size selected and the section thickness used. If thinner sections were used or small or larger nodules were included in the sample data set, then the results would be different. There were a number of limitations to our study. First, we included only pulmonary nodules of a limited size range. The purpose of our study was to assess CT section thickness in therapeutic response assessment, and we chose nodule sizes that are acceptable according to the rules governing Response Evaluation Criteria in Solid Tumors and the World Health Organization, the two most commonly used trial methods. Furthermore, we did not assess other tumor sites because heterogeneity in liver lesions and the complex size and/or shape of lymph node metastases will affect measurements in other ways not related to section thickness alone. We could have chosen to evaluate other section thicknesses; however, those evaluated in our study are the most commonly used worldwide in clinical trials and are also suggested in response criteria guidelines. In conclusion, CT section thickness reconstruction will affect the measurement of tumors. The measurement of tumors is affected by the magnitude of change in Zhao et al
Radiology
the CT section thickness reconstruction, as well as by the methods of tumor measurement. Our data indicate that the largest effect of change is between 3.75 and 7.5 mm and in volumetric measurement. The smallest effect is between 7.5 and 5.0 mm and in unidimensional measurement. References 1. Oken MM, Creech RH, Tormey DC, et al. Toxicity and response criteria of the Eastern Cooperative Oncology Group. Am J Clin Oncol 1982; 5:649 – 655. 2. World Health Organization. WHO handbook for reporting results of cancer treatment. WHO offset publication no. 48. Geneva, Switzerland: World Health Organization, 1979. 3. Miller AB, Hoogstraten B, Staquet M, et al. Reporting results of cancer treatment. Cancer 1981; 47:207–214. 4. Watson JV. What does “response” in cancer chemotherapy really mean? BMJ 1981; 283:34 –37. 5. Green S, Weiss GR. Southwest Oncology Group standard response criteria, endpoint definitions and toxicity criteria. Invest New Drugs 1992; 10:239 –253. 6. Therasse P, Arbuk SG, Eisenhauer EA, et al. New guidelines to evaluate response to treatment in solid tumors. J Natl Cancer Inst 2000; 92:205–216. 7. Saini S. Radiologic measurement of tumor size in clinical trial: past, present, and future. AJR Am J Roentgenol 2001; 176: 333–334. 8. Eggli KD, Close P, Dillon PW, Umlauf M, Hopper KD. Three-dimensional quantita-
Volume 234
䡠
Number 3
9.
10.
11.
12.
13.
14. 15. 16.
17. 18.
tion of pediatric tumor bulk. Pediatr Radiol 1995; 25:1– 6. Zhao B, Reeves AP, Yankelevitz DF, Henschke CI. Three-dimensional multicriterion automatic segmentation of pulmonary nodules of helical computed tomography images. Opt Eng 1999; 38:1340 –1347. Zhao B, Kostis W, Reeves AP, Yankelevitz DF, Henschke CI. Consistent segmentation of repeat CT scans for growth assessment in pulmonary nodules: image processing. Proc SPIE 1999; 3661:1012–1018. Zhao B, Yankelevitz DF, Reeves AP, Henschke CI. Two-dimensional multi-criterion segmentation of pulmonary nodules on helical CT images. Med Phys 1999; 26:889 – 895. Judy PF, Jacobson FL, Zhao B, Israel DA, Del Frate C. CT lung nodule size: effects scanners and reconstruction filters (abstr). Radiology 2001; 221(P):162. Zhao B, Schielke LH, Coll DM, Smith RC. Accuracy and reproducibility of renal stone volume using unenhanced helical CT and an automated 3D edge detection algorithm (abstr). Radiology 2000; 217(P):223. McCulloch CE, Searle SR. Generalized, linear and mixed models. New York, NY: Wiley, 2001. Tukey JW. The problem of multiple comparisons. Mimeographed notes. Princeton University, Princeton, NJ, 1953. Levene H. Robust tests for equality of variances. In: Olkin I, ed. Contributions to probability and statistics. Palo Alto, Calif: Stanford University Press, 1960; 278 –292. Brown MB, Forsythe AB. Robust tests for the equality of variances. J Am Stat Assoc 1974; 69:364 –367. Warr D, McKinney S, Tannock I. Influence of measurement error on assessment
19.
20.
21.
22.
23.
24.
25.
26.
of response to anticancer chemotherapy: proposal for new criteria of tumor response. J Clin Oncol 1984; 2:1040 –1046. Quoix E, Wolkove N, Hanley J, et al. Problems in radiographic estimation of response to chemotherapy in small cell lung cancer. Cancer 1988; 62:489 – 493. Lavin PT, Flowerdew G. Studies in variation associated with the measurement of solid tumors. Cancer 1980; 46:1286 –1290. Hopper KD, Kasales CJ, Van Slyke MA, et al. Analysis of interobserver and intraobserver variability in CT tumor measurements. AJR Am J Roentgenol 1996; 167: 851– 854. Thiesse P, Ollivier L, Di Stefano-Louineau D, et al. Response rate accuracy in oncology trials: reasons for interobserver variability. J Clin Oncol 1997; 15:3507–3514. Schwartz LH, Ginsberg MS, DeCorato D, et al. Evaluation of tumor measurements in oncology: use of film-based and electronic techniques. J Clin Oncol 2000; 18: 2179 –2184. Winer-Muram HT, Jennings SG, Meyer CA, et al. Effects of varying CT section width on volumetric measurement of lung tumors and application of compensatory equations. Radiology 2003; 229: 184 –194. Liang Y, Meyer CA, Krishinamurthi G, Winer-Muram HT, Jennings SG. Partial volume effect on CT lung nodule volume measurement: theoretical analysis and experimental evaluation (abstr). Radiology 2002; 225(P):560. Gundersen HJ, Jensen EB. The efficiency of systematic sampling in stereology and its prediction. J Microsc 1987; 147:229 – 263.
Pulmonary Metastases: Effect of CT Section Thickness
䡠
939