| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
Technical Developments |
1 From the Departments of Radiology (Y.A., R.L.W.), Internal Medicine (Nuclear Medicine) (S.M., R.L.W.), and Otolaryngology and Head and Neck Surgery (G.T.W.), the University of Michigan Medical Center, B1G505 University Hospital, 1500 E Medical Center Dr, Ann Arbor, MI 48109-0028. From the 1996 RSNA scientific assembly. Received April 1, 1998; revision requested June 25; revision received October 2; accepted December 16. Supported in part by National Institutes of Health grants CA 52880, MO1RR 00042, CA 53172, and CA 56731, and the High Technology funding initiative of the University of Michigan Medical Center and Clinical Research Center. Address reprint requests to R.L.W. (e-mail: rwahl@umich.edu).
 |
Abstract |
|---|
 TOP Abstract IntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
Index terms: Head and neck neoplasms, diagnosis, 20.30 • Head and neck neoplasms, emission CT (ECT), 20.12163
|
Introduction |
|---|
|
TOPAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
Accurate interpretation of FDG PET images of the head and neck, however, is not necessarily straightforward. On FDG PET images typically obtained at 50–70 minutes after radiotracer injection, physiologic FDG uptake can be seen in lymphoid tissue, nasal mucosa, or salivary gland. Radioactivity in large vessels may mimic focal tumor uptake especially at earlier imaging times. Muscle uptake may be substantial in certain cases (13). Postirradiation or surgical changes may increase FDG uptake in the soft tissue, which may not be easily distinguished from uptake associated with tumor recurrence. Such physiologic FDG uptake in the head and neck can be asymmetric, particularly postoperatively, simulating a pathologic condition at FDG PET. These findings suggest that information other than the static distribution of FDG activity after tracer injection is necessary to improve accuracy of FDG PET image interpretation.
On the basis of findings in autoradiographic studies, it has been speculated that time courses of FDG uptake may have valuable information in the distinction of neoplastic from nonneoplastic tissue activities (14). FDG uptake within a tumor gradually increases over time after intravenous administration of the radioisotope. Nonneoplastic cellular activities have been shown to have a different kinetic behavior of FDG uptake from that of neoplastic tissues. Intravascular FDG activity initially increases following injection and rapidly diminishes thereafter. Thus, examination of time courses of FDG uptake may give us additional insight to distinguish tumor versus nontumor tissues on FDG PET images.
Principal components analysis (PCA) is a multivariate correlation analysis (15) that explains algebraically a variance-covariance structure of observed data sets with a few linear combinations of original variables. When applied to dynamic FDG PET image sets, PCA finds clusters of pixels that have similar kinetic behavior of FDG uptake and summarizes them into a component. This method was applied previously in dynamic FDG PET of the liver in two dimensions (16).
PCA of dynamic FDG PET images was chosen to be explored in this study for two reasons. First, tumor tissues are heterogeneous, and conventional compartment models may not describe tumor FDG uptake adequately. This requires more complex modeling that does not guarantee the validity of the model among different tumors (17). Model-independent, data-driven analysis such as PCA may have an advantage in this regard. Second, results of PCA can be easily depicted as parametric images, which are suitable for visual interpretation of FDG PET image data.
The aim of this study was to explore the use of PCA to detect recurrent head and neck cancers. Our underlying hypothesis was that use of FDG kinetics as side information or as an alternative to a static FDG distribution image would improve the accuracy of FDG PET in the distinction of tumor from nontumor tissues in the head and neck. In this initial study, fully automated PCA was implemented in three dimensions, and resultant parametric image sets were interpreted in comparison with conventional standardized uptake value (SUV) image sets.
|
Materials and Methods |
|---|
|
TOPAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
The study was conducted with the approval of the institutional review board, with the radiopharmaceutical FDG (in-house production) administered under conditions of an investigational new drug application on file with the U. S. Food and Drug Administration. Written informed consent was obtained from all patients.
Dynamic FDG PET
Dynamic FDG PET was performed with two scanners (CTI/Siemens, Knoxville, Tenn): an Ecat Exact 921 scanner in nine patients and an Ecat 931/08–12 scanner in six patients. All patients underwent FDG PET after at least 4 hours of fasting. Transmission scanning for attenuation correction was performed before emission scanning with a germanium 68 ring source. Transmission and emission images were obtained between the level of the external auditory canal and the thoracic inlet depending on clinical symptoms suspected for tumor recurrence. After intravenous administration of 370 MBq (10 mCi) of FDG, dynamic image sets were obtained between 0 and 60 minutes after injection.
To examine PCA in various imaging protocols, image data obtained with different dynamic imaging protocols were tested. In four patients, six 10-minute frames were obtained (total, six frames). In eight patients, six 10-second frames, three 20-second frames, two 90-second frames, one 5-minute frame, and five 10-minute frames were obtained (total, 17 frames). In one patient, six 10-second frames, three 20-second frames, two 90-second frames, five 5-minute frames, and three 10-minute frames were obtained (total, 19 frames). In two patients, 12 10-second frames, three 20-second frames, two 60-second frames, five 5-minute frames, and three 10-minute frames were obtained (total, 25 frames).
Images were reconstructed with a Hanning filter (cutoff frequency, 0.3 cycles per projection element), resulting in approximately 12-mm full width at half maximum in-plane resolution. The reconstructed voxel size for the Ecat 931 scanner was 4.69 mm in plane with 6.75-mm section thickness and for the Ecat Exact scanner was 4.22 mm in plane with 3.375-mm section thickness.
PCA Method
In PCA, a three-dimensional dynamic image data set was rearranged first in a data matrix (n x p, where n is the number of dynamic frames and p is the number of pixels within a frame). Pixels within a single dynamic frame were rearranged into a single column of the data matrix from the rightmost anterior pixel to the leftmost posterior pixel of the body and then from the top to the bottom sections. A variance-covariance matrix was derived from the data matrix and then standardized to a correlation matrix (p x p). PCA was used to solve eigenvalue-eigenvector pairs from the correlation matrix. This mathematic operation is described elsewhere (15). One practical difficulty was to manipulate a large correlation matrix on a workstation with a limited amount of memory and computational power since the number of pixels within a single frame p is typically greater than 5 x 105. To overcome this problem, the data matrix was transformed and transposed to a p x n matrix to allow subsequent calculation of eigenvalue-eigenvector pairs in a correlation matrix n x n (18,19). Initial estimates of principal components were then rotated to a simpler structure by means of a Varimax rotation method (20). Resultant principal component image sets consisted of elements of rotated eigenvectors, and the number of principal components was no greater than n, the number of dynamic frames.
Pixel values in resultant PCA image sets represented correlation coefficients with latent principal components, with the maximum value of 1 being most significantly correlated with the component. For each principal component, each frame of the dynamic image set was segmented with a given PCA image set by means of multiplication. A time-activity curve for each principal component was then calculated by averaging radioactivity within pixels with the highest 1% of correlation coefficients in the PCA image matrix. Along with the time-activity curves for principal components, segmented FDG image sets of the last dynamic frame (50–60 minutes after injection) in each PCA image set were used for image interpretation (hereafter, "PCA image sets") (Fig 1). The last dynamic frame was also used to create an SUV image set (21).
|
PCA and SUV Image Evaluation
Time-activity curves and PCA image sets were evaluated by consensus of a head and neck radiologist (Y.A.) and a nuclear medicine physician (S.M.) along with clinical information including locations of primary cancers, previous treatment, presenting symptoms, and pathology results. SUV image sets were also available at the time of interpretation of PCA image sets, which was performed in two steps. First, time-activity curves of principal components were reviewed in each patient. According to the previous knowledge of FDG accumulation in viable tumors (14,22,23), components with gradually increasing FDG uptake over time or gradually increasing FDG uptake followed by a relatively stable phase were considered to be tumor or tumorlike components, respectively. PCA image sets corresponding to such time-activity curves were considered tumor PCA image sets, and they were inspected visually to localize tumor or tumorlike activities. Tumor PCA image sets were compared with an SUV image set obtained in the same patient.
To compare distinction of tumor activities from physiologic nontumor activities, structures that were visible on tumor PCA image sets and SUV image sets were examined. In this comparison, activities suspicious for or similar to tumor uptake were defined first on PCA and SUV image sets. These activities were then localized to specific structures on CT or MR images or to locations of tumor recurrence confirmed pathologically. In addition, tumor-to-background contrast ratios (CR) were calculated for both tumor PCA and SUV image sets by means of the following equation: CR = |tumor activity - background activity| ÷ (tumor activity + background activity). Background activity was measured at neck structures contralateral to tumors. By definition, the minimum and maximum contrasts were 0 and 1, respectively. A nonparametric sign test was used to assess differences in contrast ratios on PCA versus SUV image sets.
|
Results |
|---|
|
TOPAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
Inspection of time-activity curves of principal components revealed 10 patients with tumor (n = 8) or tumorlike (n = 2) components (Figs 2, 3). PCA failed to detect a tumor component in one patient with adenoid cystic carcinoma of the parotid gland (Fig 4). One recurrent tumor that was not detected with an SUV image set was also not differentiated with PCA from physiologic soft-tissue activities. In contrast, no tumor component was detected in three patients who had no recurrent tumors. When the diagnostic accuracy of PCA was calculated on the basis of identification of tumor components on a patient-by-patient basis, the sensitivity and specificity, respectively, were 83% (10 of 12 patients with tumor recurrence) and 100% (no tumor component detected in three patients without recurrence).
|
|
|
The number of identifiable physiologic soft-tissue activities on SUV image sets was reduced markedly on tumor PCA image sets with minimal loss of tumor detectability (Table). Particularly, vascular activities were suppressed completely from tumor components. Other physiologic activities such as lymphoid tissues and postirradiation or surgical changes were suppressed on tumor PCA image sets, but suppression was not complete compared to that of the vascular structures.
|
|
Discussion |
|---|
|
TOPAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
The analysis of tumor FDG PET data by means of kinetic information has been explored previously. Compartmental analysis is one of the common approaches to estimate tumor biochemical parameters. However, a difficulty with this approach is "compartmental" modeling of tumor tissues, as they may be highly heterogeneous and their compartments potentially cannot be determined precisely in human subjects. The PCA is at the opposite end of the spectrum in that it does not rely on modeling; instead, it explores variance structures of observed data. Because of its exploratory nature and pixel-by-pixel approach, the current method does not require prospective placement of regions of interest, which obviates biases potentially introduced by observers. PCA distinguishes different kinetic components contributing to the same image pixel (such as blood flow vs metabolism of the tumor), and the magnitudes of contribution of different components are summarized into different component image sets. This type of parametric imaging cannot be achieved with region-of-interest analysis on conventional SUV image sets. However, resultant component image sets can be analyzed subsequently by placing regions of interest to obtain quantitative indexes (such as estimations of contrast ratios in this study). Thus, these two approaches can be used complementarily.
Other parametric imaging techniques such as influx imaging require either sequential arterial sampling to estimate an input function of radiotracer or manual placement of regions of interest within a vascular structure (2). The former technique is invasive and may not be necessarily suitable for routine clinical use. The latter technique is suitable for thoracic neoplasms, and the drawing of regions of interest in the major vascular structures in the chest, such as aorta, is relatively straightforward. In contrast, accurate placement of a region of interest within the carotid artery in the neck can be difficult on FDG PET images. PCA is, again, advantageous in this regard since it does not require either input function or manual placement of regions of interest in the vascular structures. In addition, PCA can be used in preprocessing of dynamic images to segment vascular components, which can be used subsequently to estimate input function on dynamic images. The use of PCA for preprocessing of dynamic PET images is still under investigation. The vascular structures can be discriminated from tumor uptake by simply comparing an early postinjection FDG PET image with a standard plateau-phase image. PCA not only distinguished vascular structures but also identified different physiologic components on the basis of a variance-covariance matrix of observed data. One noticeable advantage of PCA is suppression of the background random noise component, which results in improved signal-to-noise ratios for tumors.
The application of pixel-by-pixel PCA to PET is a relatively new field. Pedersen and colleagues (16) applied the technique to abdominal tumors in two dimensions (on a single PET section) to distinguish tumor activity from nontumor background activities. A similar technique was also applied to PET brain activation data to elucidate patterns of intercorrelation of regional brain functions (24). The method to deal with a large variance-covariance matrix for PET images in the PCA was investigated by Friston and colleagues (25) for PET brain activation studies by means of an iterative algorithm. We implemented a faster and more accurate analytic algorithm (18) to process large three-dimensional volumetric dynamic PET images on a common workstation. Compared with analysis of a single section (16), inclusion of the whole image data set improves the stability of separating kinetically intercorrelated pixels in the head and neck region such as vascular components. Execution time for the algorithm was not measured systematically in the current study, but the analysis was completed in 3–8 minutes depending on the number of frames and sections of dynamic PET studies. This suggests a potential routine application of such computationally extensive analysis to clinical FDG PET studies in conjunction with evaluation of conventional SUV images.
Certain limitations exist in PCA analysis. The capability of PCA to detect tumors is governed by kinetic behaviors of FDG in tumor and nontumor soft-tissue structures. For example, the cerebral cortex has FDG kinetics similar to those of tumors, and it can be clustered in the same component, although such structures can be distinguished easily by means of correlative anatomic imaging. In addition, the analysis itself does not give the physiologic importance of each component. Thus, side information such as a time-activity curve for the component or correlative anatomic imaging is essential when results of PCA analysis are interpreted. In the current implementation, a time-activity curve for each component was generated automatically by the algorithm. This was found to be extremely useful because observers could screen the results by using time-activity curves to determine if components suspicious for tumors were seen in the observed data set. PCA and any other kinetic analysis are sensitive to patient motion during dynamic imaging, which may be improved by employing a frame-to-frame image registration algorithm.
Several dynamic imaging protocols were tested empirically in this study. The maximum number of components is equal to the number of frames (ie, a 17-frame study provided 17 components). Component image sets with large eigenvalues were displayed first in the current implementation. The first few component image sets often contained tumors. The last several components often indicated noise components. However, in a patient with a small tumor seen on a dynamic PET study covering the large field of the upper mediastinum and the neck, the tumor component was identified in the sixth component. We did not observe definite correlation between the number of frames acquired during dynamic imaging and the detectability of tumors with PCA, although this evaluation is still not complete because of insufficient experience with different imaging protocols. The greater the number of frames, the more detailed an analysis could be performed. The optimal or minimal number of frames necessary for accurate PCA analysis at dynamic PET must be established in further investigations.
In conclusion, we demonstrated the feasibility of pixel-by-pixel, three-dimensional PCA application to dynamic FDG PET tumor imaging. PCA incorporates kinetic information about FDG uptake in parametric image display and interpretation, improving tissue characterization on FDG PET images. Findings in this study demonstrate improved distinction of tumors in the head and neck regions with PCA images as compared to that with conventional SUV images. Implementation of the method permits an acceptable execution time for a routine analysis. A prospective evaluation of PCA versus static FDG imaging will be of considerable interest, as well as the application of the PCA method to tumors located in other areas of the body.
|
Acknowledgments |
|---|
|
Footnotes |
|---|
Author contributions: Guarantors of integrity of entire study, Y.A., R.L.W.; study concepts, Y.A., S.M., R.L.W.; study design, Y.A., R.L.W.; definition of intellectual content, Y.A., R.L.W.; literature research, Y.A., R.L.W.; clinical studies, R.L.W., G.T.W.; data acquisition, R.L.W., G.T.W.; data analysis, Y.A., S.M.; statistical analysis, Y.A., S.M.; manuscript preparation, all authors; manuscript editing, Y.A., S.M., R.L.W.; manuscript review, R.L.W., G.T.W.
|
References |
|---|
|
TOPAbstractIntroductionMaterials and MethodsResultsDiscussionReferences |
|---|
This article has been cited by other articles:
|
|
 |
|
  V. Gregoire, K. Haustermans, X. Geets, S. Roels, and M. Lonneux PET-Based Treatment Planning in Radiotherapy: A New Standard? J. Nucl. Med., January 1, 2007; 48(1_suppl): 68S - 77S. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 F. Dammann, M. Horger, M. Mueller-Berg, H. Schlemmer, C. Claussen, J. Hoffman, S. Eschmann, and R. Bares Rational Diagnosis of Squamous Cell Carcinoma of the Head and Neck Region: Comparative Evaluation of CT, MRI, and 18FDG PET Am. J. Roentgenol., April 1, 2005; 184(4): 1326 - 1331. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 J.-F. Daisne, T. Duprez, B. Weynand, M. Lonneux, M. Hamoir, H. Reychler, and V. Gregoire Tumor Volume in Pharyngolaryngeal Squamous Cell Carcinoma: Comparison at CT, MR Imaging, and FDG PET and Validation with Surgical Specimen Radiology, October 1, 2004; 233(1): 93 - 100. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
| RADIOLOGY | RADIOGRAPHICS | RSNA JOURNALS ONLINE |
