| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
1 All authors: Division of Cardiothoracic Imaging and the Image Processing Laboratory, Emory University Hospital and Clinic, 1324 Clifton Rd. N.E., Ste. E-118, Atlanta, GA 30322.
Received December 3, 2001;
accepted after revision January 22, 2002.
Address correspondence to J. F. Gruden.
Abstract
 Top Abstract IntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
MATERIALS AND METHODS. We retrospectively identified 25 patients with metastatic disease, each having from two to nine nodules that were 3-9 mm in diameter. Two senior and three junior reviewers interpreted all images on a workstation. The observers first reviewed axial images (3.75-mm collimation, 3-mm reconstruction interval, multidetector acquisition) in cine and sequential fashion and recorded the size, lobe, and central or peripheral (within 1 cm of the edge of lung) location of each nodule. MIP images (10-mm slab, 8-mm interval) were then reviewed, and additional nodules detected were recorded. Final counts were established by consensus.
RESULTS. The reviewers found 122 nodules (71 peripheral, 51 central) in the 25 patients. The addition of MIP slabs significantly enhanced reviewer detection of central nodules (p < 0.001) and junior reviewer detection of peripheral nodules (p < 0.001). MIP slabs also reduced the effects of reviewer experience, particularly for peripheral nodules.
CONCLUSION. MIP processing reduces the number of overlooked small nodules, particularly in the central lung. Observer nodule detection remains imperfect even when lesions are clearly depicted on images.
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
Helical CT is the technique of choice in nodule detection [3, 4]. However, failure of both helical CT to depict and the reviewer to detect small (< 6-7 mm in diameter) lesions is well known. Overall sensitivity is only 47-69% for such small nodules in clinical practice, although specific CT techniques and viewing conditions vary widely in the published studies [5,6,7,8]. Lung cancer screening failures also occur; half of the carcinomas detected on helical CT in one screening program were present in retrospect on a prior screening examination [9]. Both technical parameters (nodule depiction) and interobserver variability (nodule detection) are important determinants of overall CT sensitivity, although the relative importance of each factor is not well understood.
Multidetector CT enables simultaneous increased z-axis coverage and thinner slice collimation in comparison to single-detector helical CT. Thin sections enhance resolution and decrease volume averaging from slice to slice and should result in more accurate depiction of small nodules. However, two major factors limit observer detection of such nodules: large numbers of axial images are generated, which leads to reviewer fatigue during interpretation; and, on individual thin sections, nodules mimic normal vessels in cross-section and vice versa—particularly in the central lung zones [10, 11]. These considerations limit realization of the true potential of multidetector CT in the detection of nodules in the lung.
New computer-based image processing tools can facilitate the full use of large volumetric multidetector CT data sets. One example is maximum-intensity-projection (MIP) imaging, first described by Napel and colleagues [12]. This tool uses ray projection techniques through a stack of preselected axial images; the highest density object encountered by the ray traversing the stack is projected onto the final image (Fig. 1). MIP processing has several advantages: vascular structures appear as clearly tubular and branching structures rather than as discreet nodules; the MIP slab preserves the resolution inherent to the axial images from which it is created; and image numbers are markedly reduced in comparison to the axial image set.
|
The goal of this project was to assess the incremental effect of the addition of MIP processing to axial image review on the ability of reviewers of various levels of experience to detect small lung nodules on multidetector CT data sets.
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
Multidetector CT studies were performed on a single machine (LightSpeed; General Electric Medical Systems), and the single breath-hold acquisition parameters were identical in all patients (high speed mode; pitch, 6; table speed, 15 mm/sec; collimation, 3.75 mm; 140 kVp; 180-210 mA). Images were reconstructed at 3-mm intervals with a high frequency reconstruction (lung) algorithm and a field of view of 30-34 cm and displayed at standard window settings (width, 1200 H; level, -700 H).
A radiologist who did not participate in the study reconstructed the MIP image slabs on all study patients from the initial axial images on the workstation using a standard protocol (slab thickness, 10 mm; reconstruction interval, 8 mm; window width, -1200 H; window level, -700 H). Our prior investigations determined this protocol to be superior to thicker (30-mm) or thinner (5-mm) MIP slabs for the purpose of nodule depiction. Less than 30 sec per patient was required to produce the image slabs, which were saved and stored on the workstation along with the conventional axial images for subsequent study review.
Five reviewers independently interpreted both the axial and MIP images at two separate sessions. Observers were grouped into junior (one third-year radiology resident and two fourth-year residents) and senior (two cardiothoracic imaging faculty members) reviewers. From clinical experience, all were familiar with the normal appearance of multidetector CT and MIP images, workstation manipulation, and the cine display of numerous images. The junior resident reviewer had 2 months of previous focused training in cardiothoracic imaging that included CT and MIP image analysis; the senior residents had 4 months' training.
Reviewers were instructed to identify all nodules between 3 and 9 mm in diameter and were aware that all study patients had from two to nine lesions. Smaller nodules were ignored for the purposes of this study. Nodules were defined as any rounded opacity in the lung not attributable to a pulmonary vessel. Pleural and fissural nodules and elliptic, oblong, or linear opacities were excluded, as were areas of ground-glass attenuation.
The initial interpreting session involved the two faculty reviewers and was performed 3 weeks after the initial case selection. Images marked on the workstation as those of study patients were reviewed in random order. The axial images were initially interpreted in cine mode at two frames per second, first in the craniocaudal and then in the caudocranial direction. This initial cine overview is standard at our institution and mimics our clinical practice. The individual axial images were then reviewed sequentially for an additional 3 min. Image movement was initially controlled at the workstation by the senior radiologist to ensure complete image review during the allotted time and to establish relatively standardized viewing conditions. The initial controlled image display lasted between 90 sec and 2 min; during this time, either reviewer could ask to stop the motion or request to scroll up or down without restriction. The individual reviewers were free to independently control image movement during the remainder of the allotted time. This total time period was arbitrarily selected to approximate clinical practice and was not considered limiting by either reviewer in any case reviewed. Observers recorded all nodules, including lobar location (the lingula was considered a separate lobe), location on the image (central or peripheral), and the image number on which each nodule was best identified. Peripheral nodules were defined as those within 1 cm of the edge of the image (the outer lung) including the space along the cardiomediastinal silhouette; all other nodules were considered central.
Immediately after the axial images were reviewed, MIP images were reviewed for a total of 1 min with sequential image movement again initially controlled by the senior radiologist. Either observer could request scroll up or down functions or control image movement as in the axial image review session. Observers individually recorded any nodules not previously noted on axial image review, the image number and lobar location, and the central or peripheral location in the lung. Cine MIP review was not performed.
The second interpreting session, which involved the three junior reviewers, followed the same procedure as the first session, and was performed under the control and supervision of the senior radiologist. The senior and junior reviewers worked in separate case review sessions to minimize the number of individuals viewing the single monitor.
All nodules seen only on MIP images by any observer were subsequently confirmed and accurately localized to specific lobes on the basis of rereview of the axial images by all observers at two separate final consensus sessions. Final nodule counts (the gold standard for study purposes) were ascertained at these consensus sessions performed at 1 week and at 1 month after the conclusion of the initial reviews. All observers agreed on final nodule counts. Four patients were excluded from the final analysis because more than nine nodules were found to be present after review of the axial and MIP images. Thus, 25 patients formed the final study group for which data are presented.
The number of reviewers (n = 5) multiplied by the total number of nodules (n = 122) represented potential data points for statistical analysis. The chisquare test was performed using an alpha level of 0.5 to assess the effect of MIP images on total nodule count and on the central nodule and peripheral nodule counts. We analyzed the overall performance of reviewers and performed separate analysis for the junior and senior reviewers to determine the effects of reviewer experience on the potential benefits of MIP processing. The average number of images was calculated for both the axial and MIP image sets.
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
Table 1 shows the total number of nodules missed by all reviewers before and after the addition of the MIP processing. The incremental benefit of MIP slabs was significant (p < 0.001) for both central and peripheral nodules and for all nodules considered together (Table 1). The incremental benefit of MIP slabs was also significant (p < 0.001) for both senior and junior reviewers for the total number and, in the case of the junior reviewers, for both central and peripheral nodules (p < 0.001, Tables 2 and 3). A measurable but not statistically significant increase in the number of peripheral nodules detected by the senior reviewers was seen after the addition of MIP slabs (Table 2).
|
|
|
The senior reviewers performed better than the junior reviewers in the detection of nodules of all types before the use of MIP processing (p < 0.001). MIP slabs reduced the effect of observer experience: for peripheral nodules, the significance of the difference in the number of lesions missed by senior and junior reviewers was eliminated. In addition, the addition of MIP slabs to the axial images enhanced the performance of the junior reviewers to levels not statistically different from that of the senior reviewers interpreting axial images only (Tables 2 and 3).
The individual reviewers missed 25-48 (20-39%) of the 122 nodules, including 17-31 (33-61%) of the 51 nodules in a central location and eight to 17 (11-24%) of 71 peripheral lesions on the initial axial image review. The number of missed nodules was fairly uniform across all reviewers, but we found a statistically significant difference in the number of nodules missed by one of the senior observers compared with the other reviewers for central, peripheral, and all nodules (p < 0.001). Other individual differences were present, although none reached statistical significance. All observers missed significantly more central than peripheral nodules (p < 0.001).
The addition of the MIP slabs reduced the number of missed nodules for all individual reviewers: individuals missed 11-22 (9-18%) of the 122 nodules, five to 16 (10-31%) of the 51 central lesions, and four to seven (5-10%) of the 71 peripheral nodules (Figs. 2A,2B,2C,2D and 3A,3B,3C,3D). The addition of MIP slabs eliminated interobserver variability in the detection of peripheral nodules, and the addition of MIP processing enhanced the performance of all reviewers to levels nearly identical to that of the best axial image reviewer for central and for all nodules.
|
|
|
|
|
|
|
|
Most of the nodules missed by at least one reviewer were also overlooked by at least one other reviewer. Lesions in the medial right lower lobe (including the azygoesophageal recess) and in the medial upper lobes (particularly on the left) were problematic for multiple reviewers (Figs. 4A,4B,4C,4D,4E and 5A,5B,5C,5D,5E,5F).
|
|
|
|
|
|
|
|
|
|
|
The average number of axial images reviewed was 91 (range, 81-101). The average number of MIP images reviewed was 25 per patient slab (range, 22-28).
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
Nodule Depiction
Nodule depiction is largely a function of CT technique. Volumetric helical
CT acquisitions are superior to axial CT when all other factors are constant,
largely because the technique eliminates the respiratory misregistration and
reduces the motion artifacts of helical CT. The magnitude of this superiority
is inversely related to lesion diameter for nodules of less than 1 cm
[3]. Helical acquisitions also
enable the creation of retrospective overlapping image reconstructions that
markedly reduce the problem of slice-to-slice volume averaging. Several
clinical and experimental studies have shown the benefit of overlapping
reconstructions, particularly in the depiction of nodules of diameter equal to
or less than the slice collimation
[7,
14,15,16].
Narrow collimation (slice thickness) is intuitively desirable for depicting small lesions. One study found 3 mm to be optimal [10]. However, the trade-off between collimation and anatomic coverage area along the z-axis places a practical lower limit on collimation at 5-7 mm for single-detector helical CT of the thorax. Multidetector CT eliminates these limitations. Rapid volumetric scanning of the entire chest can now be performed with collimation as low as 2.0-2.5 mm in most patients during a single breathhold. Such acquisitions generate hundreds of images for review, which is impractical in the clinical setting and thus has limited full use of the benefits of multidetector CT.
Our multidetector CT protocol used a 3.75-mm collimation with a reconstruction interval of 3 mm and was based on earlier reports of optimal collimation for nodule detection [10]. We intentionally focused on the ability of reviewers to detect nodules 3 mm or larger. Lesions of this diameter are readily depicted using our scanning parameters; all reviewers concurred as to final nodule counts in all patients. Smaller lesions could occasionally be identified, but we did not wish to assess lesions that were not optimally depicted or were so small that their presence or absence could not be confirmed with certainty. We cannot therefore address the overall sensitivity of our multidetector CT protocol for the detection of all nodules. Rather, we chose to assess reviewer detection of clearly depicted lesions with and without the use of supplemental MIP image processing. We cannot comment on the sensitivity of multidetector CT in nodule detection; all nodules in our patients were clearly present on the available images. The limitations were in the area of observer detection.
Nodule Detection
Nodule detection is dependent on properties intrinsic to the lesion (size,
density, location), specific viewing conditions (cine or film-based, image
size, the number of images to be reviewed), and inherent characteristics of
human reviewers (including observer experience). Nearly all of these factors
are ultimately dependent on the basic target-to-background signal ratio: that
is, the contrast between the lesion and the surrounding lung architecture.
Observers tend to miss small lesions that are depicted within a complex
background [17].
The location of the nodules in the lung affects the likelihood of reviewer detection. Naidich et al. [11], using artificially placed nodules on images obtained with conventional CT, showed that perihilar or central nodules, particularly those less than 5 mm in diameter, were "extremely difficult" to identify. Croiselle et al. [5] also noted significant differences in observer detection rates for central and peripheral nodules in both the simulated and the clinical setting. Our study confirms these observations; our reviewers uniformly missed more central than peripheral lesions, with 33-61% of clearly depicted central lesions initially overlooked. Most of the missed peripheral nodules were also medial in location, in the paraspinal or paramediastinal portions of the lung (Figs. 3C, 3D, and 4A,4B,4C,4D,4E). These could also be considered central if our definitions were modified. We agree with prior investigators that difficulties in the detection of central or medial lesions reflect the nature of the complex background present in the normal central lung, with numerous normal vessels present in these areas, many of which may mimic nodules or vice versa—particularly on thinly collimated images [10, 11, 18, 19].
Viewing conditions also influence nodule detection. Prior studies have shown the benefit of cine-based review over film-based review and the potential impact of image size on nodule detection [17,18,19], and we used cine-based axial image review in our study. Cine review helps distinguish obliquely oriented vessels coursing through the scanning plane from nodules. Nodules of similar size and density to adjacent vessels can nonetheless be missed, as shown in our study and in others [11, 19]. We used both fixed and variable-frame-rate cine-based review of images at the largest possible size on the clinical workstation; image size was identical for the axial and MIP studies. However, we chose not to use true cine viewing of the MIP slabs in an effort to bias the study toward the axial image sets; it is possible that we therefore underestimated the incremental benefit of MIP slabs.
Large numbers of images may also negatively affect viewing conditions, particularly if multiple studies must be interpreted. Our average number of axial images per patient was 91, and although we did not address performance differences for reviewers at the beginning or end of the interpreting sessions, the total study involved the assessment of more than 2000 axial images. Reviewing this number of images is difficult even with cine-based display and may be an area for future inquiry.
Finally, in our study as in others, observer experience was clearly of importance in accurate nodule detection [14]. Comparison between senior and junior reviewers revealed a significant difference in the number of missed nodules of all types, and we found some variability among individual observers as well.
Effect of MIP Images
Coakley et al. [20] first
reported the use of MIP imaging applied to 3-mm collimation axial images
obtained with a dual-slice scanner in a study of simulated calcified nodules
in a canine model. Those authors found that the use of MIP increased the odds
of nodule detection by a factor of 2.18 compared with conventional axial
images having a collimation of 5 mm
[20]. Other investigators have
reported the use of 10-mm collimation MIP slabs created from contiguous 1-mm
axial images in the assessment of infiltrative micronodular disease, but with
single-detector helical CT; these investigations were limited to the study of
only a small portion of the chest
[21,
22]. Despite this technical
limitation, MIP images enhanced detection and characterization of tiny nodules
because of improved depiction of pulmonary vessels and enhanced anatomic
orientation [22].
MIP imaging preserves structures having high attenuation in the final image and displays the continuity of such structures that run obliquely through a slab of observer-selected thickness [12]. Small nodules present on the source axial images are therefore projected onto the final image along with branching blood vessels (Fig. 1). Multidetector CT enables creation of images of narrow collimation that can be reconstructed in overlapping fashion through the entire chest, and this essentially provides a contiguous "high resolution" chest CT examination. As a result, MIP images can now be constructed, and even overlapped, through the entire chest. This manipulation cannot be done with axial or even single-detector helical CT. These images take less than 30 sec to construct on our standard workstation, and creating them is now part of our routine clinical practice.
We found a significant incremental benefit in using MIP images as an adjunct to axial image review for all observers. The percentage of missed lesions fell sharply, from 20-39% to 9-18% for individual reviewers. The impact was particularly important in the identification of central nodules for reviewers, irrespective of experience: missed lesions fell from 33-61% to 10-31%. This outcome is not surprising, given the known difficulty of distinguishing vessels from nodules of similar size within a complex background. The detection of peripheral nodules was also improved, although this finding reached significance only for junior reviewers, and most of the peripheral lesions detected only with MIP processing were actually medial in location.
The incremental benefit of MIP images was marked, despite our intentional bias toward axial image review; much less time (1 vs 3 min) was allowed for MIP review in our study design. Cine review was also not performed in MIP image analysis.
The addition of MIP images also reduced, although it did not entirely eliminate, the differential effects of reviewer experience on nodule detection. The performance of all reviewers was essentially identical with regard to peripheral nodule detection with MIP as an adjunct. Junior reviewers using MIP slabs in addition to axial image review performed as well as the senior reviewers using axial images alone. This finding is of potential clinical importance.
Although MIP processing reduces image numbers to 20-30 in each patient, it is currently a tool of "second intent" [22]. The axial scans still must be reviewed until the technique is further assessed. All lesions visible on MIP images must be confirmed and accurately localized on the axial images because depth information is absent on the MIP images: visible nodules may lie on any of the individual slabs comprising the final image (Fig. 4A,4B,4C,4D,4E). In addition, at the upper and lower extremes of the lung parenchyma (apices and bases), projection of soft-tissue density from the base of the neck and the diaphragm, respectively, can result in the obscuration of small nodules on the MIP slab. This loss of information occurs when structures with high attenuation in the confines of the selected slab are projected onto the final image and obscure other smaller, high-attenuation objects in the stack (Fig. 6A,6B,6C,6D). Theoretically, nodules that project over blood vessels may also be obscured on MIP images, but this limitation is more than offset by the ability to more confidently identify nodules that are clearly separate from adjacent vessels, especially compared with identification on routine axial thin-section images [22]. MIP images should not, however, substitute for axial image review.
|
|
|
|
Although MIP images currently add to the total number of images to be reviewed in each patient, it is possible that initial MIP image review will allow the reviewer to perform a more focused and rapid assessment of the large axial image set, with particular attention to the apices and bases of the lung. Much of the difficulty in interpreting large numbers of axial images lies in the assessment of the central lung zones, where potential nodules are superimposed on a complex background. MIP images make this task simpler and more accurate, and thus allow a more directed review of the axial images. We now view the MIP images before viewing the full axial image set.
MIP slabs can be proscribed with any slab thickness or reconstruction interval. On the basis of our preliminary experience with the technique, we chose a 10-mm slab with an 8-mm (20%) overlap—similar to what has been used in prior investigations [20,21,22]. It is possible that MIP images will be of higher quality and greater value if they are constructed from thinner source axial images (1.0- to 2.5-mm collimation) or from images with a greater degree of image overlap. Variable slab thickness, with thinner slabs at the apices and bases than in the central portion of the chest, may also be desirable to minimize marginal information loss.
MIP images can be constructed in any plane. We chose the axial for several reasons: first, because section thickness exceeds the in-plane pixel width, the voxels are elongated in the z-axis, and MIP projections in planes other than the axial would lead to pixels that are no longer isotropic, which impairs resolution [12]. Second, axial MIP images preserve conventional viewing relationships used by radiologists. However, further study of various MIP protocols is warranted.
Limitations
The design of our study was somewhat artificial. However, the initial
slow-cine display is standard at our institution, and the time allotted for
review of the axial images (3 min after the cine review) was not considered
limiting by any of the reviewers. We purposely controlled the viewing
conditions to focus on the incremental benefit of MIP images rather than on
potential variability in individual performance related to differences in
other parameters. Theoretically, if observers viewed the axial images for an
infinite period of time, the incremental benefit of MIP could be reduced or
eliminated, but we doubt this to be the case, given the nature of nodule
detection. In addition, the time allotted parallels and may well exceed that
devoted to lung-window image review in a busy clinical practice.
Study of the specific characteristics of each nodule missed by each individual observer and the changes in these parameters over time for junior and senior reviewers may be of potential interest. We did not collect this data in sufficient detail, because it was not the focus of our study. Further study of the details of individual performance over time may be warranted, particularly given the recent interest in the quality control of nodule detection (screening) programs.
In conclusion, we found that MIP images enhance reviewer nodule detection of multidetector CT—depicted lesions. This is particularly true for central nodules and for lesions similar in size and appearance to normal parenchymal vessels. MIP slabs also eliminate some of the interobserver variability in nodule detection and enable the performance of junior reviewers to more closely approximate that of experienced senior reviewers. The impact of multidetector CT and MIP imaging on the detection of lesions smaller than 3 mm in diameter remains to be determined. Further research is needed to optimize multidetector CT acquisition and reconstruction protocols and MIP slab parameters to best use this promising technique.
On the basis of current knowledge, we believe that MIP images should be part of lung cancer screening programs and assessment of patients with suspected metastatic disease to the lung. Nodule detection remains imperfect even with the addition of MIP processing, and further research in computer-assisted diagnosis (CAD) and other methods of image processing is essential [5, 23].
|
TopAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
CiteULike 
Complore 
Connotea 
Del.icio.us 
Digg 
Reddit 
Technorati What's this?
This article has been cited by other articles:
|
|
 |
|
  K. J. Nichols, M. B. Tomas, G. G. Tronco, J. N. Rini, B. D. Kunjummen, K. S. Heller, L. A. Sznyter, and C. J. Palestro Preoperative Parathyroid Scintigraphic Lesion Localization: Accuracy of Various Types of Readings Radiology, July 1, 2008; 248(1): 221 - 232. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 M Bellomi, G Veronesi, C Rampinelli, S Ferretti, E De Fiori, and P Maisonneuve Evolution of lung nodules <=5 mm detected with low-dose CT in asymptomatic smokers Br. J. Radiol., September 1, 2007; 80(957): 708 - 712. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
P. Peloschek, J. Sailer, M. Weber, C. J. Herold, M. Prokop, and C. Schaefer-Prokop Pulmonary Nodules: Sensitivity of Maximum Intensity Projection versus That of Volume Rendering of 3D Multidetector CT Data Radiology, May 1, 2007; 243(2): 561 - 569. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 C. Beigelman-Aubry, C. Hill, A. Guibal, J. Savatovsky, and P. A. Grenier Multi-Detector Row CT and Postprocessing Techniques in the Assessment of Diffuse Lung Disease RadioGraphics, November 1, 2005; 25(6): 1639 - 1652. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 J. K. Leader, T. E. Warfel, C. R. Fuhrman, S. K. Golla, J. L. Weissfeld, R. S. Avila, W. D. Turner, and B. Zheng Pulmonary Nodule Detection with Low-Dose CT of the Lung: Agreement Among Radiologists Am. J. Roentgenol., October 1, 2005; 185(4): 973 - 978. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
U. J. Schoepf and P. Costello CT Angiography for Diagnosis of Pulmonary Embolism: State of the Art Radiology, February 1, 2004; 230(2): 329 - 337. [Abstract] [Full Text] [PDF]
|
