| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
Nuclear Medicine |
1 From the Division of Nuclear Medicine, Department of Radiology, Massachusetts General Hospital and Harvard Medical School, 55 Fruit St, Boston, MA 02114. Received August 9, 2001; revision requested September 7; revision received October 22; accepted December 11. Address correspondence to the author.
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMATERIALS AND METHODS RESULTSDISCUSSIONREFERENCES |
|---|
MATERIALS AND METHODS: Radionuclide perfusion images were obtained in 120 patients with normal chest radiographs who also underwent angiographic PA pressure measurement within 3 days of the radionuclide study. An artificial neural network (ANN) was constructed from several image parameters describing statistical and boundary characteristics of the perfusion images. With use of a leave-one-out cross-validation technique, this method was used to predict the PA systolic pressure in cases on which the ANN had not been trained. A Pearson correlation coefficient was determined between the predicted and measured PA systolic pressures.
RESULTS: ANN predictions correlated with measured pulmonary systolic pressures (r = 0.846, P < .001). The accuracy of the predictions was not influenced by the presence of pulmonary embolism. None of the 51 patients with predicted PA pressures of less than 29 mm Hg had pulmonary hypertension at angiography. All 13 patients with predicted PA pressures greater than 48 mm Hg had pulmonary hypertension at angiography.
CONCLUSION: Meaningful information regarding PA pressure can be derived from noninvasive radionuclide perfusion scanning. The use of image analysis in concert with artificial intelligence methods helps to reveal physiologic information not readily apparent at visual image inspection.
© RSNA, 2002
Index terms: Computers, neural network • Hypertension, pulmonary, 564.783 • Lung, radionuclide studies, 60.12171 • Radionuclide imaging, 60.12171
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODS RESULTSDISCUSSIONREFERENCES |
|---|
|
MATERIALS AND METHODS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODS RESULTSDISCUSSIONREFERENCES |
|---|
Data were available for all patients regarding the size, chronicity, and location of the emboli, as shown at angiography. The patient data in this study were collected and used in a manner determined by the hospital institutional review board to meet criteria for formal exemption from review and patient informed consent.
Chest Radiography
Half of the chest radiographs obtained at the time of the radionuclide study were posteroanterior and lateral examinations, with the other half obtained in the anteroposterior semiupright position. Normal interpretations of these images specified the absence of pleural effusion, parenchymal consolidation, lung resection, pneumothorax, and metallic artifacts such as pacemakers. Subsegmental or platelike atelectasis was not considered to be abnormal for purposes of the study because this condition does not ordinarily interfere with clinical image interpretation.
V-P Imaging
All V-P examinations were performed with a gamma camera (Orbiter; Siemens Medical Systems, Hoffman Estates, Ill). V-P scanning was performed with the patient in the supine position and breathing 370–740 MBq of xenon 133 through a closed delivery system. Wash-in, equilibrium, and washout images were obtained in the posterior projection. Although ventilation images were obtained for the clinical purpose of excluding pulmonary embolism, they were not used for the purposes of this study. Perfusion images were obtained after intravenous administration of 150 MBq of technetium 99m macroaggregated albumin with the patient in the supine position and with use of a 256 x 256 matrix size. All acquisitions totaled 1,000 K counts each. Only the anterior and posterior perfusion images were quantified for this study.
Pulmonary Angiography
Pulmonary angiography was performed in accordance with the PIOPED (Prospective Investigation of Pulmonary Embolism Diagnosis) guidelines (1). The studies were all interpreted by experienced angiographers at this institution. Separate injections were performed in each lung with magnified oblique views of the lung bases. Pressures were measured with a transducer before administration of contrast material.
The documented results indicated presence or absence of acute pulmonary embolism, size and location of the embolism, presence or absence of signs suggesting chronic embolism, and PA systolic pressure before administration of contrast material.
Image Parameters
The perfusion images were evaluated with a user-independent, whole-lung region-of-interest (ROI) method. First, an edge-detection program was used to separate the right and left lungs on the posterior and anterior perfusion images. The algorithm adjusted the lower gray-scale cutoff on each image to the first setting that permitted separation of the two lungs as independent ROIs (ie, with count-free pixels entirely separating the two lungs). A rectangular ROI was placed by the author around each resultant lung.
Each ROI thus represented the entire right or left lung. Quantification was performed on the posterior and anterior perfusion images by using an image analysis program (Optimas 6; Media Cybernetics, Bothell, Wash). The mean count density per pixel in the ROI, the area (in pixels) of the ROI, and the SD of the counts per pixel in the ROI were obtained. From these data, parameters expressing the ratio of the SD to the mean counts per pixel were obtained for the right and left lungs on the posterior perfusion views.
Heterogeneity of perfusion in the peripheral portions of the lungs was analyzed in the following manner. Four perfusion images were analyzed from each patient’s study: left anterior, left posterior, right anterior, and right posterior perfusion images. The lungs were isolated by means of placing a fixed rectangular ROI surrounding either the right or left lung. A posterization function was then applied to convert the gray-scale image to three, four, five, or six equal gray levels. The lowest level was set at 0; thus, 33%, 25%, 20%, or 17% of the lowest count density pixels were removed from the image. This procedure created a new lung outline in which the edges of the original whole-lung ROI were altered proportionally to the number of low-count pixels that constituted that edge on the original image (Fig 1). These four new images thus resulted from the removal of the four different proportions of the low-count pixels. The fractal dimension of the boundaries of each of these four modified whole-lung ROIs was calculated by using the fractal dimension plug-in (Image Processing Toolkit 3.0; Reindeer Games, Asheville, NC) for Adobe Photoshop (Adobe Systems, San Jose, Calif).
|
Neural networks can identify the predictive interactions between inputs and limit the influence of irrelevant inputs. The inclusion of too many inputs, however, given the relatively small number of training cases, may lead to a situation in which the network fails to generalize and instead "memorizes" irrelevant attributes of the training cases. Given the relatively small number of training cases, it was important to limit the number of inputs to those most likely to be relevant. The perfusion contour of the anterior views of the lungs is more strongly influenced by the heart and mediastinal structures than is that of the posterior views. This normally greater contour irregularity in the anterior views suggested they should be analyzed differently than the more regularly contoured posterior views. We addressed this by limiting, a priori, analysis of anterior perfusion images to those in which either 17% or 20% of low-count pixels were removed before fractal dimension calculation. This somewhat arbitrary decision limited the number of inputs to the neural network and avoided the inclusion of data primarily influenced by the mediastinal contour.
The 16 parameters selected for input to the artificial neural networks (ANNs) were the mean count density per pixel and the SD of the counts per pixel for posterior perfusion images of the left and right lungs (four inputs); the fractal dimensions of the right and left lung posterior perfusion images after removal of 17%, 20%, 25%, and 33% of the lowest count pixels from the image (eight inputs); and the fractal dimensions of the right and left lung anterior perfusion images after removal of 17% and 20% of the lowest count pixels from the image (four inputs).
Artificial Neural Networks
All ANNs were constructed with a computer program (Neuroshell2; Ward Systems Group, Frederick, Md). A back propagation network was used to predict PA pressure. The ANN included 16 inputs and seven hidden nodes and was trained until the mean error of the network decreased to below a uniform, predetermined level. A single output node indicated the predicted systolic PA pressure, in millimeters of mercury. A logistic activation function was applied to the hidden layer. The ANNs were tested by using the leave-one-out cross-validation method (3), in which all cases but one were used to train the ANN, which was then applied to the single excluded case. This procedure was repeated for each case, such that each case was left out only once (ie, used as the test case for an ANN trained on the remaining cases). Calculations were performed on a dual processor 933-MHz Pentium III personal computer using 640-MB PC133 RAM, constructed by the author.
Statistical Analysis
Standard correlation analysis (SPSS Base 9.0; SPSS, Chicago, Ill) was performed. Pearson correlation coefficients were obtained with a two-tailed P value to determine whether the prediction of the network performed significantly better than chance. Significance was defined as a two-tailed P value of less than .05. Other data were expressed as means ± standard errors of the mean. Comparisons of means were performed by using the t test.
|
RESULTS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODS RESULTSDISCUSSIONREFERENCES |
|---|
Figure 2 shows the individual correlation coefficients of the input parameters with the measured PA systolic pressures at angiography. The highest individual correlation coefficient was 0.432 (P < .01); and the lowest, 0.070 (not significant). Figure 3 shows the scatterplot of ANN predictions of PA pressure versus the pressures measured at angiography. The Pearson correlation coefficient was 0.846 (P < .001). The standard error of the estimate was 7.09 mm Hg. The correlation coefficient between ANN prediction and angiographic measurement in patients with pulmonary embolism was 0.848. In patients without pulmonary embolism, the correlation coefficient was 0.845.
|
|
|
|
DISCUSSION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODS RESULTSDISCUSSIONREFERENCES |
|---|
Previous studies have attempted to predict the degree of arterial obstruction at angiography from semiquantitative estimates of film density on perfusion images (10). These authors found a hyperbolic relationship between the extent of pulmonary vascular obstruction estimated from the lung scan and the PA pressure in acute pulmonary embolism. Higher total peripheral resistances were seen with chronic thromboembolism than with acute embolism for a given degree of vascular obstruction (10). This difference was attributed to small-vessel disease in the patients with chronic embolism, which was inapparent at subjective evaluation of the perfusion lung scan. Changes in the peripheral small vessels may thus be an important parameter in determining the PA pressure.
Patients with primary pulmonary hypertension and chronic thromboembolic pulmonary hypertension are known to show medial hypertrophy and intimal proliferation in the small distal arteries of the lung (10). This small-vessel component may alter the appearance of the lung periphery on perfusion scans. Although the scans of most patients with primary pulmonary hypertension lack segmental perfusion defects, the images obtained in some of these patients show a nonspecific "patchiness" that may be related to small-vessel disease (11). This subtle, but physiologically important small-vessel component could contribute to the lung scan’s qualitative underestimation of the severity of pulmonary hypertension in chronic thromboembolic disease (12).
These prior studies suggest the pursuit of a quantitative parameter that describes perfusion heterogeneity, particularly in the periphery of the lung. One way of defining peripheral heterogeneity is to progressively subtract low-count pixels from the perfusion image and analyze the resultant lung contour. Scans with less heterogeneity at the lung periphery will produce more regular outlines and maintain this regularity as the lung border is progressively altered by deleting low-count pixels from the image. Quantitative characterization of the lung periphery can then be performed by using fractal techniques.
Fractals are structures that possess self-similarity in space or time over a range of measurement resolution. The structure of a fractal object appears similar at both macroscopic and microscopic scales. The lung is known to possess fractal characteristics (13) in that the branching of the tracheobronchial tree shows a similar structure in both its larger central and smaller peripheral branchings. The PA tree follows a similar pattern. Some workers have suggested that the fractal characteristics of biologic objects correlate with their degree of adaptability and general health (14). It seems reasonable to suppose that the fractal characteristics of the PA tree might be altered in pulmonary hypertension. For this reason and also because of the sensitivity of fractal analysis to image boundary characteristics, fractal analysis was a logical method to use in this study. Previous workers have used the concepts of fractal analysis to analyze both radionuclide (15–18) and other radiologic images (19).
Fractals are often characterized by a number known as the fractal dimension. This index describes the space-filling properties of an object—in a sense, its "wiggliness." Mathematically, it represents the self-similarity of an object across different measurement scales or the degree to which the object possesses fractal characteristics. There are several standard methods to calculate the fractal dimension of an object (20). I chose a method of fractal dimension calculation that is suitable for the examination of an object’s boundary, the Minkowski dilation method. This parameter, as applied to the perfusion images described herein, shows different values, depending on the heterogeneity of perfusion at the lung periphery. For instance, removing 20% of the lowest count density pixels will produce an image of the lung that differs from the initial image, depending on the number and location of these low-count pixels. If the lowest count pixels are irregularly concentrated in the periphery of the lung, as might occur in patients with pulmonary hypertension, this removal will have a relatively larger effect on the resultant lung boundary and thus on the fractal dimension of this image.
The fractal dimension was calculated at different thresholds to better characterize perfusion at the periphery of the lungs. Sample images resulting from these manipulations are shown in Figure 1. The original posterior perfusion images are shown at the left, with progressive alterations proceeding to the right (each produced by deleting low-count density pixels from the original image and thus defining a new perfusion contour). Note that the contours appear more "wiggly" on the images obtained in the patient with severe pulmonary hypertension (top image set) than they do on the images obtained in the patient with a normal PA pressure (bottom image set). The fractal dimension generally increases as a greater percentage of background counts is subtracted from the image. For a given amount of background subtraction, however, the fractal dimension is greater in those patients with high PA systolic pressures, reflecting the greater heterogeneity of perfusion in these patients.
In addition to the fractal dimension, I included global spatial parameters of the type used in previous studies involving the use of neural networks (8,9)—specifically, mean pixel count density and variation in density between pixels on a given projection. These indexes provide a measure of global perfusion heterogeneity. It is important to note that all of the parameters selected for input into the ANN avoid subjective interventions (such as freehand ROI selection) that could bias the results and limit generalization of the findings. All ROIs were computer-defined whole-lung ROIs from which automatically computed indexes were subsequently obtained.
These several parameters were used as inputs to an ANN. The resultant network successfully estimated the pulmonary pressures observed at angiography. The Pearson correlation coefficient of the ANN predictions with measured PA systolic pressures was 0.846 (P < .001), as shown in Figure 3. The standard error of the estimate was 7.09 mm Hg. An ANN that used only the mean and SD count data (without including fractal information) showed a correlation coefficient of only 0.418, thus accounting for only 17% of the variability in PA pressure. This emphasizes the importance of the fractal dimension parameter in the prediction since its inclusion increased the correlation coefficient to 0.846. The analysis procedure could thus account for 72% of the total variability in PA pressure. The accuracy of the predictions was not affected by the presence or absence of pulmonary embolism.
The Table shows the ANN predictions in cases of mild, moderate, and severe pulmonary hypertension as defined at angiography (21). There were no patients with pulmonary hypertension among the 51 with an ANN-predicted PA systolic pressure of 29 mm Hg or less. Similarly, when the ANN predicted a pressure of 48 mm Hg or greater, all 13 patients so classified had angiographically elevated PA systolic pressures.
It may be argued that radionuclide images do not faithfully reproduce the fractal characteristics of an anatomic object. The fractal properties of an object may be altered by scatter, attenuation, noise, and collimator effects. Although from a purely practical perspective the fractal dimension effectively quantifies the spatial heterogeneity of the recorded image in an objective manner, networks of the type described herein may be dependent on gamma camera characteristics and acquisition protocols. This could limit the portability of the algorithm and necessitate retraining under the new operating conditions. I believe that the clinical application of this method, at least at this stage of development, is most appropriate to ancillary screening in that it yields information that is in addition to that traditionally obtained from the perfusion lung scan. Further refinements in predictive accuracy might make the method applicable to diagnosis and treatment follow-up.
In conclusion, a meaningful correlation exists between measures of image heterogeneity on perfusion lung scans and measured PA pressure at angiography. The use of the fractal dimension, suggested by the fractal nature of the lungs in vivo and the importance of small-vessel changes in pulmonary hypertension, is one approach to noninvasively estimate PA pressure. ANN predictions might be improved by using higher count perfusion images with larger matrix sizes and thus better defining peripheral perfusion heterogeneity and optimizing the fractal dimension calculation. Future work will determine whether such additional costs in imaging time would be repaid in the coin of useful physiologic data.
|
FOOTNOTES |
|---|
Author contribution: Guarantor of integrity of entire study, J.A.S.
|
REFERENCES |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODS RESULTSDISCUSSIONREFERENCES |
|---|
This article has been cited by other articles:
|
|
 |
|
  U. Tateishi, H. Uno, K. Yonemori, M. Satake, M. Takeuchi, and Y. Arai Prediction of Lung Adenocarcinoma Without Vessel Invasion: A CT Scan Volumetric Analysis Chest, November 1, 2005; 128(5): 3276 - 3283. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
| RADIOLOGY | RADIOGRAPHICS | RSNA JOURNALS ONLINE |
