DOI:10.2214/AJR.06.0843
AJR 2007; 188:1239-1245
© American Roentgen Ray Society
Multiprojection Correlation Imaging for Improved Detection of Pulmonary Nodules
Ehsan Samei1,2,3,
Stanton A. Stebbins1,
James T. Dobbins, III1,3 and
Joseph Y. Lo1,3
1 Duke Advanced Imaging Laboratories, Department of Radiology, Duke University
Medical Center, 2424 Erwin Rd., Suite 302, Durham, NC 27705.
2 Department of Physics, Duke University Medical Center, Durham, NC.
3 Department of Biomedcial Engineering, Duke University Medical Center, Durham,
NC.
Received June 29, 2006;
accepted after revision November 21, 2006.
Address correspondence to E. Samei.
Supported in part by grants R01CA109074 and R01CA80490 from the National
Institutes of Health.
There is a research agreement between GE Healthcare and the laboratory of
J. T. Dobbins, III, involving research grant funding and intellectual property
agreements. None of the activities represented in this article are a part of
that agreement, although an X-ray detector on loan by GE Healthcare was used
in the experiments described in the article. None of the coauthors of this
article stand to benefit financially from the work presented.
Abstract
OBJECTIVE. The purpose of this study was the development and
preliminary evaluation of multiprojection correlation imaging with 3D
computer-aided detection (CAD) on chest radiographs for cost- and
dose-effective improvement of early detection of pulmonary nodules.
SUBJECTS AND METHODS. Digital chest radiographs of 10 configurations
of a chest phantom and of seven human subjects were acquired in multiple
angular projections with an acquisition time of 11 seconds (single
breath-hold) and total exposure comparable with that of a posteroanterior
chest radiograph. An initial 2D CAD algorithm with two difference-of-gaussians
filters and multilevel thresholds was developed with an independent database
of 44 single-view chest radiographs with confirmed lesions. This 2D CAD
algorithm was used on each projection image to find likely suspect nodules.
The CAD outputs were reconstructed in 3D, reinforcing signals associated with
true nodules while simultaneously decreasing false-positive findings produced
by overlapping anatomic features. The performance of correlation imaging was
tested on two to 15 projection images.
RESULTS. Optimum performance of correlation imaging was attained
when nine projection images were used. Compared with conventional, single-view
CAD, correlation imaging decreased as much as 79% the frequency of
false-positive findings in phantom cases at a sensitivity level of 65%. The
corresponding reduction in false-positive findings in the cases of human
subjects was 78%.
CONCLUSION. Although limited by a relatively simple CAD
implementation and a small number of cases, the findings suggest that
correlation imaging performs substantially better than single-view CAD and may
greatly enhance identification of subtle solitary pulmonary nodules on chest
radiographs.
Keywords: cancer • chest • computer-aided detection • lung disease • physics • radiologic physics
Introduction
The mortality rate associated with lung cancer is higher than that
of any other form of cancer
[1]. Because lung cancer is
frequently detected only at late stages, the prognosis often is poor. The key
to improving early detection is to identify small, solitary lung nodules
before the cancer has metastasized to the lymphatic system and other organs.
Detection of small lung nodules on chest radiographs, however, is often
hindered by perceptual errors compounded by anatomic noise whereby normal
anatomic structures, which have a camouflaging effect, obscure and mimic
disease [2]. To address this
challenge, computer-aided detection (CAD) techniques have been developed to
search for nodule-like opacities on chest radiographs
[3-5].
By facilitating thorough search of images, CAD works as a second reviewer to
assist the radiologist and thereby reduces perceptual errors inherent in
visual examination of chest radiographs. As is visual detection by human
observers, however, CAD of a single chest radiograph is hampered by anatomic
noise, leading to relatively low specificity for an acceptable level of
sensitivity. One key to improving the performance of CAD in nodule detection
is to reduce the number false-positive findings due to anatomic noise.
Correlation imaging is a new multiprojection imaging technique designed to
improve the performance of CAD algorithms in the detection of pulmonary
nodules on chest radiographs. Correlation imaging entails the use of multiple
low-dose digital chest radiographs acquired at different projection angles and
processed with a CAD algorithm. Geometric correlation with CAD results is used
to differentiate true nodules from anatomic background, improving the
detection of lung nodules while reducing the number of false-positive
findings. The radiation dose in correlation imaging is comparable with that
delivered for a standard posteroanterior chest radiograph. The purpose of this
study was to introduce the technique, ascertain the optimum number of
projection images for correlation imaging, and evaluate use of the technique
in examinations of a phantom and of human subjects.
Subjects and Methods
Image Data
Image data containing nodule opacities were acquired from a chest phantom
and from human subjects with institutional review board approval. All image
data were acquired with a prototype X-ray system developed at our institution
for a parallel research project on chest tomosynthesis
[6]. The system consisted of a
flat-panel detector (similar to that in the XQ/i system, GE Healthcare), which
produced 2,048 x 2,048 pixel, 14-bit chest images with a 0.2-mm pixel
pitch. All data were acquired in raw format (bad-pixel and gain calibration
applied) at a 183-cm source-to-image distance with a standard 12:1 antiscatter
grid.
The phantom images were acquired with an anthropomorphic chest phantom
(Humanoid, RSD) positioned in the standard posteroanterior configuration. Two
polytetrafluoroethylene (Teflon, Du-Pont) nodule phantoms were placed on the
posterior surface. These nodules were designed to accurately duplicate the
radiographic appearance of subtle tissue-equivalent lung nodules 8-10 mm in
diameter [7]. The nodules were
repositioned in 10 configurations to produce 10 phantom cases. For each case,
a 120-kVp, 5-mAs beam setting
[8] was used to acquire 71
projection images spaced evenly over a vertical angular range of 20° in 11
seconds. Figure 1 is a
schematic of the setup. To determine the exact location of the nodules on the
images, the chest phantom also was imaged without nodules to provide a digital
subtraction mask.
View larger version (11K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 1 —Diagram shows configuration used to acquire projection images in
correlation imaging. X-ray tube moves precisely along vertical axis to acquire
projection images from required angle ( ). Displacement of lesion in
each projection image is defined by angle, source-to-image distance (S), and
distance (d) between nodule (or nodule phantom) and detector.
|
|
The human subject data set consisted of images of four men and three women
(age range, 48-69 years; median age, 56 years) with CT-confirmed lung nodules.
A total of 18 nodules with a size range of 1-21 mm (average size, 8.1 mm;
median size, 7.0 mm) were present on the images. The images were acquired with
the aforedescribed method with the exceptions that the effective tube current
was adjusted for the thickness of each subject and that 61 (rather than 71)
projection images were acquired within the 20° angular range. The
procedure, equivalent to that used for the tomosynthesis research at our
institution, involved total X-ray exposure equivalent to that of a lateral
chest radiograph [6]. Because
the correlation imaging method used in this project relied on a subset of only
two to 17 projection images, the correlation imaging technique involved 3-25%
of the dose of a lateral chest radiograph.
In addition to the multiprojection phantom and human subject images
acquired in this study, a set of 44 retrospective digitized posteroanterior
chest images were used as test cases for selection of the appropriate
parameters for the CAD algorithm. With institutional review board approval,
these images were selected from an existing, anonymous data set of 300 images
acquired previously [9,
10]. Human subjects whose
radiographs were included in this data set had regions on a chest radiograph
suspicious for pulmonary nodules that were further evaluated with fluoroscopy
or CT [10]. Only images
containing nodules 5-15 mm in diameter were used to represent the most
clinically important sizes targeted by our technique. Nodules smaller than 5
mm are unlikely to be detected on projection images with any computational
method or by radiologists, whereas nodules larger than 15 mm are unlikely to
be overlooked by a radiologist.
Single-View CAD
We developed a single-view 2D CAD algorithm for the front end of the
correlation imaging method. Figure
2 is a flowchart of the algorithm. In step 1, each projection
image was scaled down to 512 x 512 pixels. In step 2, each projection
image was processed with difference-of-gaussians (DOG) filters
[11]. The smaller of the two
gaussian SDs typically is chosen to match a corresponding size of the
gaussian-like nodules [7]. The
larger of the two SDs is picked to generate a blurry estimate of anatomic
background structures to be removed. A sample of the DOG-filtered output is
shown in Figure 3.
View larger version (126K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 3 —Image shows sample difference-of-gaussians (DOG) filter output from
anthropomorphic phantom projection. Bright areas represent areas of greatest
similarity between input image and DOG filter output.
|
|
In this study, we used two separately applied DOG filters (as opposed to
the move conventional one DOG filter) to more fully encompass the range of
sizes (5-15 mm) of the nodules of interest. We determined the initial DOG
parameters by examining the histogram of the distribution representing the
diameters of the true nodules. The parameters were then empirically optimized
to yield the highest free-response receiver operating characteristic (FROC)
curve [12] on the set of 44
posteroanterior chest radiographs. The qualitative examination of the FROC
curves with only one DOG filter indicated that a DOG filter of 2-4 (i.e., DOG
composed of 2-pixel-wide and 4-pixel-wide gaussian filters) produces the
highest sensitivity at a given specificity level. With the first DOG parameter
kept constant at 2-4, the introduction of a second DOG filter with variable
parameters indicated an optimum FROC profile for 5-10 parameters in terms of
the most desirable compromise between sensitivity and specificity.
In step 3, a contour map was generated from each filtered output that
provided the likelihood that a nodule was present in a given portion of the
image. Twenty intensity thresholds were applied over the entire intensity
range of the DOG-filtered output. A useful analogy to this procedure is the
process of lowering a water level over a topography of geographic features and
hills (i.e., nodules). As the water level goes down, islands begin to emerge
above the surface of the water and grow larger. At each threshold, islands, or
nodules, that were too small, too large, or too noncircular (determined by a
ratio of perimeter to area) to represent the characteristics of true solitary
lung nodules were eliminated. Size and circularity values defining these
thresholds were chosen with a small-grid search of eight randomly selected
posteroanterior chest images. After the elimination process, the surviving
shapes from each threshold were added together to form a gray-scale contour
map. The contour maps corresponding to both DOG filter outputs were added
together to produce the final contour map for the projection.
Correlation Imaging Technique
In our implementation of correlation imaging, CAD results from multiple
projections were combined to reduce the effect of anatomic noise on nodule
detection. In this method, a shift-and-add algorithm
[6] was applied to the 2D CAD
probability distributions. This process produced a 3D probability distribution
data set that effectively reinforced signals that correlated geometrically. To
limit the complexity of the correlation imaging algorithm, the number of
projections used for the 3D reconstruction was limited to 17 for the phantom
and 15 for the human subjects. To simplify the display and evaluation of the
results, the 3D data set was summed in the z-direction, producing a
2D contour map comparable with the posteroanterior chest radiograph and
containing information from all the processed projections. In step 4 of
correlation imaging, manual lung segmentation of the posteroanterior
projection was applied to limit the presentation area to the lung region and
to produce the final output of the algorithm
(Fig. 4).
View larger version (44K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 4 —60-year-old man with pulmonary nodules. Final prethresholding
contour map is produced with correlation imaging algorithm generated by
shift-and-add tomosynthesis reconstruction of contour maps produced from each
projection as input, summation of all slices to produce 2D image, and
application of manual lung field segmentation. Bright areas represent areas
where nodule is likely to be found.
|
|
Performance Evaluation
To evaluate the performance of the approach, an FROC curve was generated by
thresholding of the contour map of the final output of correlation imaging and
comparison of the results with a truth file (step 5). The truth file was
defined as a binary mask of circular areas with specified values for the
centroids and diameters, which were determined by means of image subtraction
for the phantom or by CT data for the human subjects. If a nodule on the 2D
contour map overlapped any of the true nodule areas, a true-positive finding
was registered. All other regions that did not overlap were counted as
false-positive findings.
Results
As implemented, correlation imaging proved to be a robust technique leading
to a notable reduction in false-positive findings. Compared with conventional,
single-view CAD, correlation imaging decreased the false-positive findings in
phantom cases as much as 79% at a sensitivity level of 65%. The corresponding
reduction in false-positive findings in human subject cases was 78%. A
representative case is illustrated in Figure
5A,
5B. The circles in
Figure 5B represent
true-positive findings (actual nodule locations), and the red shapes represent
possible nodules. Although the detection algorithm operates in 3D, results are
projected on the posteroanterior image to provide context for the nodule
locations.
View larger version (92K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 5B —Anthropomorphic phantom. Sample postthresholding output for
correlation imaging algorithm. Red regions represent possible nodules detected
with correlation imaging. Circles indicate nodules in truth file. Possible
nodules that intersect circles are counted as true-positive findings, and
those that do not are counted as false-positive findings.
|
|
A key correlation imaging parameter investigated was the number of
projection images used. Figure
6A shows the FROC curves that show the compromises between
sensitivity and false-positive rate for different numbers of images used.
Qualitative examination of the graph shows there was little shift in the FROC
curves for correlation imaging with either two or three images, which produced
essentially the same FROC curves as traditional 2D CAD in the posteroanterior
projection. As the number of images used in correlation imaging increased,
however, the FROC curve shifted upward and to the left, indicating increasing
sensitivity and specificity. For example, at a fixed sensitivity of 88%,
correlation imaging yielded 40 false-positive findings with two projections
but 14 such findings with 17 projections. Corresponding numbers at 65%
sensitivity were 11 and three false-positive findings.
Figure 6B shows the relation
between the number of false-positive findings present and the number of
projections used for correlation imaging at a fixed sensitivity level, akin to
drawing a horizontal line across Figure
6A. Figure 6B shows
a clear increase in specificity for correlation imaging in five or more
projections. Specificity increased with the number of projections used, but
there appeared to be a diminishing return beyond seven projections.
View larger version (21K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 6A —Analysis of false-positive findings for phantom. Graph shows
free-response receiver operating characteristic (FROC) results for correlation
imaging in differing numbers of anthropomorphic phantom projections along with
FROC results for 2D computer-aided detection (2D CAD) output of single
posteroanterior projection.
|
|
View larger version (9K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 6B —Analysis of false-positive findings for phantom. Graph with fixed
sensitivity level of 65% shows relation between number of false-positive
findings and number of images used in correlation imaging algorithm.
|
|
A representative correlation imaging output for a human subject is shown in
Figure 7A,
7B. Superimposed over a
posteroanterior projection to provide context, the circles represent true
nodules and the red areas possible nodules detected. The human subject cases
had many more false-positive findings than the phantom cases (e.g., for
correlation imaging with seven images at a sensitivity of 65%, approximately
50 false-positive findings for humans vs approximately three for the phantom)
because the false-positive findings correlated with some of the normal
nodule-like anatomic features. FROC results for differing numbers of
projections are shown in Figure
8A. There was a clear increase in sensitivity and specificity for
correlation imaging with five or more projections.
Figure 8B shows the relation
between the number of projections used for correlation imaging and
false-positive findings for a given sensitivity threshold. Increasing the
number of projections used beyond three tended to increase specificity, but
for the human subject cases, performance continued to improve for larger
numbers of projections up to the maximum of 15 images.
View larger version (103K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 7B —60-year-old man with pulmonary nodules. Output for correlation
imaging algorithm at threshold of 66% sensitivity. Red regions represent
possible nodules detected with correlation imaging. Circles represent nodules
in truth file. Possible nodules that intersect circles are counted as
true-positive findings, and those that do not are counted as false-positive
findings.
|
|
View larger version (20K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 8A —Analysis of false-positive findings for human subject. Graph shows
free-response receiver operating characteristic (FROC) results for correlation
imaging with differing numbers of human subject projections along with FROC
for 2D computer-aided detection (2D CAD) output of single posteroanterior
view.
|
|
View larger version (9K):
[in this window]
[in a new window]
[as a PowerPoint slide]
|
Fig. 8B —Analysis of false-positive findings for human subject. Graph with
fixed sensitivity level of 65% shows relation between number of false-positive
findings and number of images used in correlation imaging algorithm.
|
|
The overall results suggest that using the maximum number of images
produces the best results. The return, however, diminishes for more than five
to seven images. This finding suggests seven images is perhaps the most
dose-effective number for use in multiprojection imaging. With seven
projections in the best scenario based on phantom data, the current
correlation imaging technique at 65% sensitivity yielded 3.4 false-positive
findings per case. This figure was based on our preliminary implementation and
is subject to improvement as the technique is further refined. In addition,
according to the particular implementation of correlation imaging in this
study on the basis of a constant dose per projection, the use of each
additional image implies a corresponding incremental increase in radiation
dose to the patient. Future implementation based on a constant target total
dose fractionated among multiple projections may yield different results.
Discussion
Although some results [13]
have suggested that chest radiography can be a reliable screening tool for
lung cancer, other results
[14,
15] have shown that
radiologists frequently miss subtle lung nodules that are discovered after a
diagnosis is made. In at least one trial
[16], investigators found no
statistically significant effect on lung cancer mortality. This limitation of
chest radiography can be attributed to one of its inherent imaging
characteristics: A standard posteroanterior radiograph reduces the 3D anatomy
of the chest to a 2D shadow. Thus, certain anatomic features may overlap and
hide abnormal features of interest, such as solitary pulmonary nodules. This
overlap of anatomic features also can produce shapes that mimic the appearance
of nodules. When a projection radiograph of the chest is taken from a
different angle, however, it is less likely that the same anatomic features
would overlap so as to obscure the same true nodules or to produce false
nodules in the same location as in the posteroanterior radiograph. A true
nodule is likely to persist across different projections, albeit in slightly
different locations because of the effects of parallax.
Correlation imaging is a new technique in which information from multiple
projections is used to reduce the influence of anatomic noise and thus
increase the sensitivity and specificity of computer-aided diagnosis with
X-ray exposure comparable with that of a standard chest radiograph. The
hardware needed for correlation imaging will likely be modestly more expensive
than that used for standard digital radiographic systems. The technique
requires a motorized gantry for translating and pivoting the X-ray tube in
synchronized motion with a high-speed X-ray pulse and digital detector data
acquisition.
Correlation imaging builds on an earlier algorithm, biplane correlation
imaging, first presented by Samei et al.
[8]. In biplane correlation
imaging, the locations of suspicious regions are compared in two projection
images, and the known parallax of geometric features are used to differentiate
between true- and false-positive findings. Biplane correlation imaging has
been shown to greatly increase specificity at the expense of a small decrease
in sensitivity [8]. Biplane
correlation imaging was designed, however, for use with only two projections
rather than the larger number of projections used in correlation imaging. In
our experiment, the gain in specificity for correlation imaging was comparable
with that reported for biplane correlation imaging. In biplane correlation
imaging, a 20% loss in sensitivity yielded a 94% increase in specificity
compared with single-view CAD
[8]. In correlation imaging
with seven phantom projections, for example, a 70% increase in specificity was
achieved with no loss in sensitivity. Correlation imaging, however, does not
appear to be similarly penalized by losses in sensitivity for increased
specificity because nodules that do not correlate in one or more of the
projections are not automatically eliminated, as they are in biplane
correlation imaging. Correlation imaging shifts FROC curves (Figs.
6A and
8A) upward and to the left,
indicating increases in both sensitivity and specificity.
The imaging system described is a one-of-a-kind investigational device, and
this study was one of the first attempts at development of automated detection
algorithms for these images. As such, no direct comparison with other
algorithms was possible. An indirect comparison, however, can be made with the
only commercially available chest radiographic CAD system (RapidScreen
RS-2000, Deus Technologies). That system operates with 87 nodule features,
heuristic decision rules, an artificial neural network, and fuzzy logic in its
algorithm. Trained on 1,000 T1 cancer radiographs and 10,000 cancer-free
radiographs and tested on 80 T1 radiographs of cancer patients with nodules
7-30 mm in diameter and 160 radiographs of patients without cancer, the system
depicted 66% of the cancerous nodules with an average of 5.3 false-positive
findings per image (Deus Technologies, U. S. Food and Drug Administration
application, 2001). When the operating point for correlation imaging was set
to a comparable 65% sensitivity, correlation imaging yielded three
false-positive findings per case with 17 phantom projections and 25
false-positive findings with 15 human subject projections.
Although they are encouraging, the results of our study are limited in
several respects. First, a relatively simple CAD algorithm was used without
extensive feature extraction and selection to reduce the frequency of
false-positive findings. This approach was chosen to ensure maximum
sensitivity in the selection of nodules for use as input to the 3D correlation
component of correlation imaging. Second, the preliminary findings of the
study were limited by the small number of cases. As new human subject cases
become available, more thorough evaluation may become possible. Third, the
results were limited by the initial reconstruction method used to correlate
the CAD output from the different projections. A different reconstruction
algorithm, perhaps a variant of filtered back-projection not pursued because
of the additional optimization needed with a small number of cases, may be of
more utility for correlation imaging.
The fourth limitation arising from the aim to simplify the performance
evaluation of correlation imaging was that the 3D probability matrix was
flattened into two dimensions. Three-dimensional evaluation would naturally be
preferred, taking into account the depth, shape, and 3D probability
distribution of the probable nodules. Fifth, although optimization of the two
DOG filters in the algorithm was relatively rigorous, other parameters, such
as nodule size or circularity thresholds and the angular distribution of the
projections, were left relatively unexplored. These parameters, although
reasonably chosen, should be optimized more exhaustively in future
implementations. Finally, and most importantly, because the investigation was
based on projection images from existing tomosynthesis scans, we compared the
correlation imaging performance from different numbers of projections for a
total exposure that linearly increased with the number of projection images
used. The exposure ranged from approximately 16% to 100% of the total exposure
of a single posteroanterior chest radiograph. Future implementation of
correlation imaging, however, may divorce the effect of exposure from the
number of projections used. Such an implementation would require a separate
experiment, the results of which may differ from those of this study.
Despite the limitations, which would have to be overcome in future
implementations, the current correlation imaging algorithm yields improved
performance over single-view CAD. This finding indicates the potential of this
technique for greatly enhancing identification of solitary pulmonary nodules
in chest radiography.
Acknowledgments
We thank Nariman Majdi-Nasab, Carey Floyd, Amarpreet Chawla, and Georgia
Tourassi of Duke University for their assistance with this study.
References
- American Cancer Society. Cancer facts and figures
2005. Atlanta, GA: American Cancer Society,2005
- Samei E, Flynn MJ, Eyler WR. Detection of subtle lung nodules:
relative influence of quantum and anatomical noise on chest radiographs.
Radiology 1999;213
: 727-734[Abstract/Free Full Text]
- Doi K. Current status and future potential of computer-aided
diagnosis in medical imaging. Br J Radiol2005; 78[spec no 1]:S3
-S19[Abstract/Free Full Text]
- Baydush AH, Floyd CE Jr. Spatially varying bayesian image
estimation. Acad Radiol 1996;3
: 129-136[CrossRef][Medline]
- Lo SC, Freedman MT, Lin JS, Mun SK. Automatic lung nodule detection
using profile matching and back-propagation neural network techniques.
J Digit Imaging 1993;6
: 48-54[Medline]
- Dobbins JT III, Godfrey DJ. Digital x-ray tomosynthesis: current
state of the art and clinical potential. Phys Med Biol2003; 48:R65
-R106[CrossRef][Medline]
- Samei E, Flynn MJ, Eyler WR. Simulation of subtle lung nodules in
projection chest radiography. Radiology1997; 202:117
-124[Abstract/Free Full Text]
- Samei E, Catarious DM Jr, Baydush AH, Floyd CE Jr, Vargas-Voracek
R. Bi-plane correlation imaging for improved detection of lung nodules.
Proc SPIE 2003;5030
: 284-297[CrossRef]
- Floyd CE Jr, Baydush AH, Lo JY, Bowsher JE, Ravin CE. Bayesian
restoration of chest radiographs: scatter compensation with improved signal to
noise ratio. Invest Radiol 1994;29
: 904-910[CrossRef][Medline]
- Drayer JA, Vittitoe NF, Vargas-Voracek R, Baydush AH, Ravin CE,
Floyd CE Jr. Characteristics of regions suspicious for pulmonary nodules at
chest radiography. Acad Radiol 1998;5
: 613-619[CrossRef][Medline]
- Zheng B, Chang YH, Gur D. Computerized detection of masses in
digitized mammograms using single-image segmentation and a multilayer
topographic feature analysis. Acad Radiol1995; 2:959
-966[CrossRef][Medline]
- Arenson RL, Chakraborty DP, Seshadri SB, Kundel HL. The digital
imaging workstation. Radiology 1990;176
: 303-315[Abstract/Free Full Text]
- Shimizu N, Ando A, Teramoto S, Moritani Y, Nishii K. Outcome of
patients with lung cancer detected via mass screening as compared to those
presenting with symptoms. J Surg Oncol1992; 50:7
-11[Medline]
- Quekel LG, Kessels AG, Goei R, van Engelshoven JM. Miss rate of
lung cancer on the chest radiograph in clinical practice.
Chest 1999; 115:720
-724[CrossRef][Medline]
- Heelan RT, Flehinger BJ, Melamed MR, et al. Non-small-cell lung
cancer: results of the New York screening program.
Radiology 1984;151
: 289-293[Abstract/Free Full Text]
- Muhm JR, Miller WE, Fontana RS, Sanderson DR, Uhlenhopp MA. Lung
cancer detected during a screening program using four-month chest radiographs.
Radiology 1983;148
: 609-615[Abstract/Free Full Text]
CiteULike 
Complore 
Connotea 
Del.icio.us 
Digg 
Reddit 
Technorati What's this?
Top
Abstract
Introduction
