Академический Документы
Профессиональный Документы
Культура Документы
251
We previously showed that high-reso- All size measurements were made on this
lution (thin sections and small field of filtered, segmented nodule image. This
view) CT imaging could be used to assess technique was used for segmentation of
nodule growth (6). In this previous study, both the synthetic and in vivo nodules.
we selected the CT image containing the We have developed an interactive
largest cross-sectional area of the nodule graphical user interface, or GUI, that per-
and determined the nodule area for each mits the radiologist to set parameters and
point in time. Using the change in the to depict the results at each step of this
nodule area, we calculated the doubling process. The system also allows for visual
time by using an exponential growth comparisons of the nodule at all stages of
model. This provides a conservative esti- the analysis.
mate of growth. We showed that malig- To help resolve partial volume effects
nant nodules all demonstrated growth and obtain subvoxel accuracy, the data
rates consistent with malignant tumors, were resampled to an isotropic space by
Figure 1. Three-dimensional shaded-surface
whereas benign nodules did not (6). using three-dimensional linear (trilinear) images from high-resolution CT scans of two
The purpose of this study was to deter- interpolation (14). This produced an im- deformable silicone synthetic nodules. Each
mine the accuracy of high-resolution CT age having higher resolution than the pair of images shows a single nodule. The im-
volumetric measurements of small pul- original image on the basis of a 0.25-mm3 age on the left was scanned first; then, after
monary nodules to assess growth and supergrid. In the new, isotropic space, we deforming the nodule, the image on the right
was obtained. These synthetic nodules span
malignancy status over time. This ap- applied a series of 3D morphologic filters
the size range that was tested. The top nodule
proach has substantial advantages, be- to help in segmenting the nodule from has an initial maximum diameter of approxi-
cause the determination of volume is less other adjacent objects, such as vessels mately 7 mm, and the bottom nodule has an
dependent on patient positioning and and bronchi, or from the pleural surface, initial maximum diameter of approximately
section selection and allows for detection while preserving the relevant nodule 14 mm.
of asymmetric growth. characteristics.
Once the nodule was segmented, vol-
ume was determined by adding the num-
in the volume or area was made, we as-
MATERIALS AND METHODS ber of voxels contained in the resultant
sumed that the nodule growth or de-
image. Similarly, the two-dimensional
crease could be represented by the simple
The use of three-dimensional (3D) tech- (2D) area was determined by counting
exponential growth model. This model
niques for nodule volume determination the number of pixels contained in the
allows us to calculate the tumor doubling
requires contiguous high-resolution CT single image having the largest cross-sec-
time. Doubling time (DT) may be ex-
images of the entire nodule. Further- tional area (6). The same segmentation
pressed in terms of direct volume (V)
more, since our focus was on small nod- techniques were applied to all repeat im-
measurements as DTV [log 2 t]/[log
ules, it was important to achieve the ages of the nodule acquired at different
(V 2 /V 1 )], where t is the change over
highest possible resolution along the times to minimize additional segmenta-
time. With the use of 2D diameter esti-
scan axis by using the smallest pitch pos- tion errors; for purposes of growth assess-
mates (D), doubling time may be calcu-
sible. Thus, our image data were acquired ment, it is more important to precisely
lated as DTD [log 2 t]/[3log (D 2 /
by using a CT scanner (HiSpeed Advan- determine the relative change in nodule
D 1 )]. By using 2D area estimates ( A),
tage; GE Medical Systems, Milwaukee, volume than the absolute volume mea-
doubling time may be calculated as
Wis) in helical mode at 1:1 pitch, 140 surements. A full description of these tech-
DTA [2log 2 t]/[3log ( A 2 /A 1 )]. It
kVp, and 200 mA by using 1-mm beam niques is beyond the scope of this article.
is important to note that the doubling
collimation. The whole nodule region To compare the accuracy of the 2D and
time calculation based on 2D diameter
was scanned in a single breath hold. The 3D techniques, we determined the nod-
estimates assumes uniform growth in three
images were reconstructed at 0.5-mm in- ule diameter, D, on the basis of the nod-
dimensions of a spherical object. Since
tervals with a 9.6-cm field of view by ule area and volume. In calculating D on
nodules do not necessarily grow uni-
using the high-spatial-resolution algorithm. the basis of the 2D nodule area, we as-
formly in all dimensions, true volumetric
Using a software package (VISIONX; Cor- sumed that the nodule was a perfect cir-
determinations theoretically should pro-
nell University, Ithaca, NY; available at: cle. Similarly, to determine D on the ba-
vide more accurate information.
http://www.ee.cornell.edu/reeves/research sis of the nodule volume, we assumed
/visx/index.html) for multidimensional im- that the nodule was a perfect sphere.
Synthetic Nodule Data
age processing and computer vision, the Since we were most interested in the
radiologist (D.F.Y.) selected a region of consistency of our technique and the A series of three experiments was per-
interest containing the nodule. The im- ability to detect relative change, we cal- formed by using synthetic nodules. In
age sections containing the region of in- culated the percentage of error of each each of these, the nodules were placed on
terest were then extracted to obtain the measurement relative to the area or vol- synthetic foam having an attenuation
reconstructed 3D image volume of the ume. In this context, we defined the per- similar to that of aerated lung. Related
region of interest. This was accomplished centage of error as the mean-normalized work in measurement of synthetic nod-
by resampling the 3D image to isotropic square root of the mean-squared error rel- ules has been described (15). These exper-
(resolution is identical in all three dimen- ative to each metric. This measurement iments were designed to test the preci-
sions) space and then thresholding to ob- also can be thought of as the coefficient sion and reproducibility of our scan and
tain a 3D binary image, followed by the of variation, or SD normalized by the image processing techniques in making
application of the 3D segmentation tech- mean, of each calculated metric. volumetric determinations. The clini-
niques used to define the nodule region. Once the determination of the change cally relevant pulmonary nodules were
Volume 217 Number 1 Small Pulmonary Nodules: Volumetrically Determined Growth Rates 253
for smaller nodules. For example, for
nodules smaller than 6 mm in diameter, TABLE 3
In Vivo Nodule Diameters and Doubling Times Based on Changes in the Volume
the root-mean-square variation was within and Area Measurements
plus or minus 1.1%; in nodules larger
than 6 mm in diameter, it was within
plus or minus 0.5% (Fig 2). The greater Volume (mm3) Area (mm2) Doubling Time
Diameter (d)
percentage of error for these smaller nod- Patient No. of Days Estimate First Second First Second Final
ules again relates to the increased propor- No. between Scans (mm) Scan Scan Scan Scan Volume Area Diagnosis*
tion of partial voxel effects on the surface 1 36 6.9 106.9 135.7 36.5 36.6 104 9,700 Malignant
of the smaller nodules. The data clearly 2 20 9.3 239.8 313.8 65.9 74.1 51 78 Malignant
illustrate the inverse relationship of mea- 3 69 5.4 141.3 184.8 18.1 25.7 177 90 Malignant
surement error to nodule size, as well as 4 71 6.5 265.2 466.4 32.6 66.9 87 46 Malignant
5 33 5.5 62.5 85.3 250.1 341.2 73 49 Malignant
our ability to detect small volume differ- 6 745 3.9 89.0 166.4 11.4 28.3 826 378 Benign
ences in nodules of similar size. 7 35 7.4 70.0 70.9 280.1 283.4 2,030 135 Benign
Results from these synthetic nodule 8 35 7.2 54.6 56.3 218.5 225.3 798 532 Benign
studies suggest that the image resolution 9 84 4.1 36.2 36.2 13.0 14.9 33,700 288 Benign
10 225 4.0 41.5 37.6 12.2 11.8 1,570 2,840 Benign
is adequate following isotropic resam- 11 61 7.1 208.6 219.3 38.9 46.3 846 164 Benign
pling to determine change in volume 12 70 8.4 207.9 222.2 52.4 53.6 731 1,520 2YNC
within a 2% error. However, when the 13 306 5.8 91.5 156.2 25.6 34.1 396 494 2YNC
shape is altered substantially (Fig 1), then * 2YNC 2 years without change.
the change in volume can be determined
within a 3% error, although this can be
reduced to 1% if the CT images are recon-
structed at 0.5-mm intervals. Further-
more, 3D volumetric measurements ap- the benign nodules had relatively short
peared to be better than or comparable to area-based doubling times: 164, 288, and
2D area measurements. 378 days for patients 11, 9, and 6, respec-
tively; the doubling times based on the
volume were 846, 33,700, and 826 days,
Patient Data respectively. Some malignant nodules
The volume and area measurements had asymmetric patterns of growth iden-
obtained from in vivo nodules are given tified by using the 3D techniques but not
in Table 3 together with the time in days the previously developed 2D methods
between the two scans. Table 3 also (eg, patient 1).
shows the estimated doubling times for
each nodule separately on the basis of DISCUSSION
volume change and area change. The fi-
nal diagnosis is shown in the last col- Three-dimensional methods for estimat- Figure 2. Scatterplot illustrates percentage of
umn: malignant, benign, or 2 years of ing growth offer intrinsic advantages error of the volume measurements in 35 spher-
follow-up without change. To appreciate over conventional 2D methods. Since the ical synthetic nodules that were 311 mm in
the approximate size of each nodule, a whole nodule is used, advantage is taken diameter and reconstructed at 0.5-mm inter-
vals. The root-mean-square (RMS) and maxi-
diameter calculation based on the initial of all available data, not just a subset
mum percentage of error for nodules 3 6 mm
volume is provided. The initial diameters (section with maximum cross-sectional in diameter are 1.1% and 2.8%, respectively.
ranged from 3.9 to 9.3 mm. area). When performing the repeat scan, For nodules 6 11 mm in diameter, these are
Doubling times based on volume mea- there is no need to select matching im- 0.5% and 0.9%. The greater variability seen in
surements were substantially smaller for ages on the follow-up study (registra- the smaller nodules reflects the increased pro-
malignant nodules than for benign nod- tion), since total volume is measured, portion of partial voxels. With increasing size,
the amount of error in volume estimation be-
ules. This difference was also seen with which makes the volumetric measure-
comes progressively smaller.
area measurements, although the differ- ments independent of registration. Reg-
ence was less clear. By using volumetric istration occasionally can cause difficulty
measurements, all malignant nodules when patients are not oriented in exactly
had doubling times of less than 177 days, the same position during repeat scans as the entire volume of the nodule must be
which is considered well within the typ- they were during the original study or, obtained in a single breath hold.
ical range for malignancy (16). All of the despite the use of thin sections, the im- In a previous study (6), we were able to
benign nodules had doubling times that ages do not correspond precisely. A final demonstrate the potential of early repeat
were 396 days or greater on the basis of advantage is the ability to view the nod- scanning with use of 2D techniques. This
volumetric measurements. When the nod- ule in 3D space from any arbitrary view- was accomplished by using both syn-
ule decreased in volume over time, as in point. This provides additional visual thetic nodules and actual patient data.
patient 10, the doubling time was nega- verification concerning the accuracy of We showed that CT has the spatial reso-
tive. the segmentation process. lution necessary for detection of change
Results of area measurement were The additional challenge in 3D volu- in volume within 30 days for nodules
more variable with one malignant nod- metric determination is that motion arti- that have doubling times within 30 180
ule (patient 1) having a long doubling facts and scanning speed introduce addi- days and are larger than 5 mm in diam-
time of 9,700 days. In addition, several of tional errors in the data acquisition. Also, eter at initial presentation, and we pre-
Volume 217 Number 1 Small Pulmonary Nodules: Volumetrically Determined Growth Rates 255
techniques (volume vs area) are due to cause this preliminary study was limited Blackburn JT, Doi K, MacMahon H. Com-
differences in the growth patterns of the by the relatively small number of cases puterized detection of pulmonary nod-
ules on CT scans. RadioGraphics 1999;
nodules themselves or differences in the evaluated. It is also important to note,
19:13031311.
measuring techniques. We have given however, that we focused on pulmonary 8. Henschke CI, Miettinen OS, Yankelevitz
examples in which one or the other was nodules smaller than 10 mm, because DF, Libby D, Smith JP. Radiographic
more likely. In patients 2, 5, and 13, the other diagnostic techniques (eg, biopsy) screening for cancer: new paradigm for its
similarity of the results is striking. As we are performed more easily in nodules scientific basis. Clin Imaging 1994; 18:
accumulate additional cases and further larger than 10 mm and it is more difficult 16 20.
9. Henschke CI, McCauley DI, Yankelevitz
refine our segmentation techniques, we to justify periods of observation in these
DF, et al. Early Lung Cancer Action
presumably will be able to gain addi- cases to establish growth rates. As with all Project: overall design and findings from
tional insight. It is clear, however, on the other diagnostic tests, the accuracy of baseline screening. Lancet 1999; 354:99
basis of synthetic nodule studies, that CT volumetric measurements declines with 105.
is capable of helping make these determi- decreasing nodule size, primarily due to 10. Munden RF, Pugatch RD, Liptay MJ, Sug-
nations, assuming that segmentation tech- the increasing partial voxel averaging. arbaker DJ, Le LU. Small pulmonary le-
sions detected at CT: clinical importance.
niques for real nodules approach the re- Despite this, 3D volumetric techniques
Radiology 1997; 202:105110.
liability they do for synthetic nodules. appear to be a promising new tool for 11. Lillington GA. Management of solitary
Although we had no overlap between obtaining additional information about pulmonary nodules. Postgrad Med 1997;
benign and malignant doubling times smaller nodules. 101:145150.
based on 3D measurements, slow-grow- 12. Lillington GA. Management of the soli-
ing malignant nodules may have dou- tary pulmonary nodule. Disease-a-month
References 1991; 37:271318.
bling times suggestive of benign disease. 13. Yankelevitz DF, Henschke CI. Does 2-year
1. Friese SA, Rieber A, Fleiter T, Brambs HJ,
There is even evidence to suggest that Claussen CD. Pulmonary nodules in spi- stability imply that pulmonary nodules
these slower growing malignant nodules ral volumetric and single-slice computed are benign? AJR Am J Roentgenol 1997;
behave less aggressively than the rapidly tomography. Eur J Radiol 1994; 18:48 51. 168:325328.
growing ones (18). However, as we gain 2. Diederich S, Lenzen H, Windmann R, et 14. Jain AK. Fundamentals of digital image
al. Pulmonary nodules: experimental and processing. Englewood Cliffs, NJ: Prentice
experience, we will be able to develop a
clinical studies at low-dose CT. Radiology Hall, 1989.
scheduling function that will allow us to 1999; 213:289 298. 15. Harris KM, Adams H, Lloyd DC, Harvey
decide when to perform additional repeat 3. Chiou YS, Lure YM. Hybrid lung nodule DJ. The effect on apparent size of simu-
studies to ascertain growth and to better detection (HLND) system. Cancer Lett lated pulmonary nodules of using three
understand the probability that a partic- 1994; 77:119 126. standard CT window settings. Clin Radiol
4. Zhu X, Lee KN, Levin DL, et al. Temporal 1993; 47:241244.
ular growth rate implies either malig- image database design for outcome anal- 16. Mizuno T, Masaoka A, Ichimura H, Shi-
nancy or benignity. Prospective evalua- ysis of lung nodule. Comput Med Imag-
bata K, Tanaka H, Niwa H. Comparison of
tion of the techniques outlined in this ing Graph 1996; 20:347356.
actual survivorship after treatment with
article should be performed toward the 5. Kanazawa K, Kawata Y, Niki N, et al.
Computer-aided diagnosis for pulmonary survivorship predicted by actual tumor-
refinement of such a model, as well as the nodules based on helical CT images. volume doubling time from tumor diam-
overall method. It seems likely that for Comput Med Imaging Graph 1998; 22: eter at first observation. Cancer 1984; 53:
certain doubling times, further diagnos- 157167. 2716 2720.
6. Yankelevitz DF, Gupta R, Zhao B, Hensch- 17. Russ JC. The image processing handbook.
tic studies will be recommended.
ke CI. Small pulmonary nodules: evalua- 2nd ed. Boca Raton, Fla: CRC, 1995.
At this time, we still need to learn more tion with repeat CTpreliminary experi- 18. Spratt JS, Meyer JS, Spratt JA. Rates of
about the growth of these small nodules, ence. Radiology 1999; 212:561566. growth of human solid neoplasms: part I.
as well as our measuring techniques, be- 7. Armato SG III, Giger ML, Moran CJ, J Surg Oncol 1995; 60:137146.