November 2000, VOLUME 175
NUMBER 5

Recommend & Share

November 2000, Volume 175, Number 5

Chest Imaging

Automatic Detection and Quantification of Ground-Glass Opacities on High-Resolution CT Using Multiple Neural Networks
Comparison with a Density Mask

+ Affiliations:
1Department of Radiology, Johannes Gutenberg-University Mainz, Langenbeckstr. 1, 55131 Mainz, Germany.

2Institute of Computer Science, Johannes Gutenberg-University Mainz, Staudinger Weg 9, 55128 Mainz, Germany.

3Institut für Radiologie, Universitätsklinikum Lübeck, Ratzeburger Allee 160, 23538 Lübeck, Germany.

Citation: American Journal of Roentgenology. 2000;175: 1329-1334. 10.2214/ajr.175.5.1751329

ABSTRACT
Next section

OBJECTIVE. We compared multiple neural networks with a density mask for the automatic detection and quantification of ground-glass opacities on high-resolution CT under clinical conditions.

SUBJECTS AND METHODS. Eighty-four patients (54 men and 30 women; age range, 18-82 years; mean age, 49 years) with a total of 99 consecutive high-resolution CT scans were enrolled in the study. The neural network was designed to detect ground-glass opacities with high sensitivity and to omit air—tissue interfaces to increase specificity. The results of the neural network were compared with those of a density mask (thresholds, -750/-300 H), with a radiologist serving as the gold standard.

RESULTS. The neural network classified 6% of the total lung area as ground-glass opacities. The density mask failed to detect 1.3%, and this percentage represented the increase in sensitivity that was achieved by the neural network. The density mask identified another 17.3% of the total lung area to be ground-glass opacities that were not detected by the neural network. This area represented the increase in specificity achieved by the neural network. Related to the extent of the ground-glass opacities as classified by the radiologist, the neural network (density mask) reached a sensitivity of 99% (89%), specificity of 83% (55%), positive predictive value of 78% (18%), negative predictive value of 99% (98%), and accuracy of 89% (58%).

CONCLUSION. Automatic segmentation and quantification of ground-glass opacities on high-resolution CT by a neural network are sufficiently accurate to be implemented for the preinterpretation of images in a clinical environment; it is superior to a double-threshold density mask.

Introduction
Previous sectionNext section

Several attempts have been made to use texture analysis techniques for the automatic segmentation of CT scans [1,2,3,4]. Classic statistical methods may include the parameters of first-, second-, and third-order statistics and other composite or custom-made texture parameters. Their disadvantages include the extensive training required before automated or semiautomated segmentation, evaluation of potential usefulness of a particular parameter only after implementation, and cumbersome and time-consuming adaptation to new segmentation tasks. Conversely, neural networks are promising for a modular approach to the segmentation of CT scans. We applied self-organizing neural networks, so-called Kohonen-feature maps [5], that use the pixel information of CT scans, such as density, or first- and second-order statistics as input [6]. For the first clinical application, we chose the quantitative assessment of ground-glass opacities on high-resolution CT scans. The neural network should “preinterpret” the scans to assist the radiologist in making a diagnosis and in determining the extent of ground-glass opacities. For this screening purpose, high sensitivity and high negative predictive value are required, and quantification is essential for follow-up examinations. In this prospective study, we investigated the usefulness of the hierarchic network of multiple neural networks for the automatic detection and quantification of ground-glass opacities under clinical conditions. A currently available and easy-to-implement approach to detect ground-glass opacities is an appropriate density mask that can be used as a reference technique to evaluate the improvement obtained by the neural network. The assessment of a radiologist served as the gold standard.

Subjects and Methods
Previous sectionNext section

Eighty-four patients (54 men and 30 women; age range, 18-82 years; mean age, 49 years) with a total of 99 consecutive CT scans were prospectively enrolled in the study, which was approved by the appropriate institutional review board. Informed consent was obtained for further processing of clinical and imaging data. The patients were referred for CT for the following clinical reasons: suspected pneumonia in patients at risk (n = 73), fibrosis or alveolitis (n = 23), acute respiratory distress syndrome (n = 1), pulmonary congestion (n = 1), and postradiation pneumonitis (n = 1). Follow-up studies were examined for nine patients (4 had 2, 4 had 3, and 1 had 4 examinations performed). The prevalence of disease-related ground-glass opacity was estimated to be approximately 50%. High-resolution CT was performed on a Somatom plus S scanner (Siemens Medical Systems, Erlangen, Germany) with 137 kV, 210 mAs, a 1-mm section thickness, a 10-mm increment, an acquisition in maximum inspiration (1 sec), a reconstruction with a high-frequency algorithm (ultrahigh), and a mean field of view of 360 mm2. From all studies, three scans were selected for segmentation. Intentionally, one scan was obtained from the upper (apex to the carina), middle (carina to the lower lung veins), and lower lung fields (lower lung veins to the diaphragm). Scans with severe pulsation or respiration artifacts were excluded. The data were transferred onto a PC-based workstation (Pentium [200 MHz, 32MB of RAM]; Intel, Feldkirchen, Germany) for postprocessing. The technical details for image segmentation, development, implementation, and testing of the neural network have been published elsewhere [6]. In summary, a hierarchy of a CT scan, two receptive fields, three neural networks, two contrast parameters, and one expert rule was built. The receptive fields consisted of 5 × 5 pixels for networks 1 and 2 and of 9 × 9 pixels for network 3. Network 1 had 15 × 15 neurons and divided the image into ground-glass pixels and non-ground-glass pixels with high sensitivity by identifying the density range from -750 to -300 H without explicit textural parameters. However, network 1 could not discriminate homogeneous areas of ground-glass opacity from other contrast-rich opaque areas close to high-attenuation structures, such as bronchovascular bundles [6]. Networks 2 and 3, both with 10 × 10 neurons, were used to detect these contrast-rich opaque areas using contrast as an explicit textural parameter; the networks identified density differences between pixels as a measure of the amount of contrast in the receptive field, and large differences corresponded to increased contrast. The different sizes of the receptive fields of networks 2 and 3 were used to account for contrast in small and large environments. Therefore, contrast richness was detected as a more local and a more mid-range feature to segment (e.g., smaller and larger vessels). An equation (expert rule) combined the three networks: Ground-glass pixels = network 1 positive pixels - network 2 positive pixels - network 3 positive pixels. For the first evaluation, the neural network computed the total number of pixels classified as ground-glass opacity. The lungs were segmented semiautomatically (thresholds, -1024/ -200 H), and the percentage of lung area affected was calculated. For comparison, a density mask purely based on the predefined ground-glass opacity thresholds (-750/-300 H) was applied, and the percentage of the affected lung area was calculated. Values for agreement and disagreement were computed. On the basis of former experience, we considered the neural network superior to the density mask, and differences between the two indicated gains of the neural network in sensitivity (areas detected only by the network) and specificity (areas detected only by the density mask). For the second evaluation, an experienced chest radiologist served as the gold standard. Initially blinded to the results of the neural network and the density mask, the chest radiologist visually evaluated the presence and extent of ground-glass opacity on a workstation using different window settings. The color-coded segmentation of the neural network was then displayed on the monitor. The same radiologist visually estimated the size of the area with a false-negative or a false-positive classification of ground-glass opacity. This measurement was given as a percentage (to the nearest 5%) of the area truly showing ground-glass opacity (gold standard). Afterward, the same procedure was repeated for the density mask, and the percentages for true-positive and true-negative findings were computed. Then sensitivity, specificity, positive predictive value, negative predictive value, and accuracy were calculated on a per-scan and per-patient basis. Differences between the upper, middle, and lower lung fields were noted. The reasons for false-positive findings, such as partial volume effects, respiratory motion, or cardiac motion were recorded. The percentage of true-positive findings corresponding to dependent opacities was also noted. The Wilcoxon's rank sum test was used for statistical evaluation. A probability of error for the entire set of less than 5% was regarded as significant (p < 0.05; Bonferroni adjustment for multiple comparisons: cutoff, 0.004).

Results
Previous sectionNext section

Semiautomated segmentation of lung required approximately 2 min per scan, and automated segmentation of ground-glass opacity required approximately 5 min per scan for the neural network and approximately 10 sec for the density mask.

In the first evaluation, the neural network classified 6% of the total lung area as ground-glass opacity (Table 1). Ground-glass opacity was least frequent in the upper lung (5.1%), more frequent in the middle lung (5.1%), more frequent in the middle lung (6%), and most frequent in the lower lung (6.9%). An average of 4.7% was also detected by the density mask. An average of 1.3% was detected solely by the neural network (Fig. 1A,1B,1C) on the basis of textural properties. This percentage represents the increase in sensitivity by the neural network. The density mask classified 22% of total lung area as ground-glass opacity, and 17.3% was detected solely by the density mask without textural features characteristic of ground-glass opacity (Table 1). We considered the neural network superior to the density mask, and false-positive areas detected by the density mask were omitted; they represented the increase in specificity by the neural network (Figs. 2A,2B,2C and 3A,3B,3C) Other abnormalities with a higher increase in density, such as masses or consolidations, were not categorized as ground-glass opacity by the network (Fig. 1A,1B,1C).

TABLE 1 Lung Area Classified as Ground-Glass Opacity After Evaluation of CT Scans with Neural Network and Density Mask

figure
View larger version (206K)
Fig. 1A. 49-year-old female patient with acute respiratory distress syndrome. Source image shows extensive ground-glass opacities and consolidation.

figure
View larger version (147K)
Fig. 1B. 49-year-old female patient with acute respiratory distress syndrome. Image with segmentation by neural network shows extensive ground-glass opacities in green with concentration in core areas. Pulmonary vessels, pleura, and consolidations are clearly delineated and not categorized as ground-glass opacities.

figure
View larger version (156K)
Fig. 1C. 49-year-old female patient with acute respiratory distress syndrome. Image with segmentation by density mask shows extensive ground-glass opacities in green. False-positive segmentations are found at borders toward pleura and bronchial walls.

figure
View larger version (88K)
Fig. 2A. 43-year-old female patient with pneumonia. Source image shows ground-glass opacities in left upper and left lower lobes, indicating pneumonia. Note dependent opacities in right lower lobe and partial volume effect in right upper lobe.

figure
View larger version (104K)
Fig. 2B. 43-year-old female patient with pneumonia. Image with segmentation by neural network shows ground-glass opacities in green, indicating pneumonia, dependent opacities, and partial volume. Note false-positive segmentation of fissure on right side.

figure
View larger version (112K)
Fig. 2C. 43-year-old female patient with pneumonia. Image with segmentation by density mask shows ground-glass opacities and all peripheral air—tissue interfaces (false-positive findings) in green.

figure
View larger version (93K)
Fig. 3A. 27-year-old male patient with suspected pneumonia. Source image does not show ground-glass opacities.

figure
View larger version (105K)
Fig. 3B. 27-year-old male patient with suspected pneumonia. Image with segmentation by neural network shows small partial volume effects at pleura and vascular borders (false-positive findings) in green.

figure
View larger version (111K)
Fig. 3C. 27-year-old male patient with suspected pneumonia. Segmentation by density mask shows air—tissue interfaces, peripheral pulmonary structures, and bronchial walls (false-positive findings) in green.

In the second evaluation, the neural network had a sensitivity of 99%, specificity of 83%, positive predictive value of 78%, negative predictive value of 99%, and accuracy of 89% for the area involved with ground-glass opacity when compared with the gold standard (Tables 2 and 3). Sensitivity and negative predictive value did not vary between the upper, middle, and lower lung fields, whereas we noted a decrease in specificity, positive predictive value, and accuracy from the upper to the lower lung fields. This decrease was caused by a significant increase in false-positive findings (overall 21.4%) from the upper lung fields (11.2%) to the middle (21.9%) and lower lung fields (27%) (Table 2). The false-positive findings were caused by partial volume effects of bronchovascular bundles, lung veins, pericardium, or pleura and pulsation artifacts. Of the true-positive findings, 9.6% corresponded to dependent opacities. All scans with ground-glass opacities were identified by the neural network. With regard to the lung area, false-negative findings averaged 1.1% with a maximum of approximately 5% on images with extensive ground-glass opacity.

TABLE 2 Evaluation of Ground-Glass Opacity Revealed by Neural Network

TABLE 3 Characteristics for Neural Network and Density Mask in Segmentation of Ground-Glass Opacities

The density mask was inferior to the neural network with a sensitivity of 89%, specificity of 55%, positive predictive value of 18%, negative predictive value of 98%, and accuracy of 58% (Table 3). The high number of false-positive findings was caused by the fact that the density mask outlined almost every air—tissue interface (bronchovascular bundles, interlobular septa, and pleura).

The neural network was also used for follow-up examinations. Resolution of ground-glass opacities was found in three patients in whom fever had disappeared. Despite persistent fevers, five other patients no longer had large areas of ground-glass opacity (>1% of total lung area) on any CT scans. In these patients, nonspecific ground-glass opacity was no longer thought to be related to the cause of the fever. A patient with alveolitis showed stable disease without a change in ground-glass opacities; this finding correlated with clinical findings.

Discussion
Previous sectionNext section

A self-organizing neural network was implemented for the automated segmentation and quantification of ground-glass opacities on high-resolution CT scans of the lung. Because of high sensitivity and high negative predictive value, it can be used in the clinical environment for the preinterpretation of high-resolution CT scans. When compared with a density mask, the neural network showed a marked improvement in specificity, positive predictive value, and accuracy.

Until now, applications of artificial neural networks in chest imaging concentrated on the detection and evaluation of pulmonary nodules and interstitial disease based on conventional chest radiography [7,8,9,10,11] and on the diagnosis of acute pulmonary embolism based on ventilation-perfusion scintigraphy [12, 13]. In most older studies, neural networks were trained with expert knowledge. For example, the radiomorphology of pulmonary nodules and the corresponding probabilities of malignancy were used as input [7]. The neural network gives the accurate or the most likely diagnosis [14], 15]. Only recent studies used digital data of chest radiography [8,9,10] or of CT [4] as input for the automatic detection of pulmonary nodules. To our knowledge, digital CT data have not been used as direct input for self-organizing neural networks to detect interstitial lung disease. The main reasons for this approach are flexibility and self-organization. These properties are advantageous in comparison with classic texture analysis (Bayesian analysis, classification of particular feature vectors, or cluster analysis) [16,17,18], which has been applied for the segmentation of CT scans with various success rates [1, 16, 19]. Because these systems are dependent on particular trainers, users, and CT scanners, their use is not widespread. Conversely, self-organizing neural networks do not require an adequate and complete set of training patterns [20], which would be extremely difficult to obtain for an extensive and amorphous texture like ground-glass opacity. These difficulties have already been stated as motivation for the application of neural networks [8]. Because of self-organization, minor adaptive changes are necessary only for the evaluation of scans acquired by different scanners.

The density range of ground-glass opacity has not been defined. From our experience and from reports in the literature, we selected values from -750 to -300 H. Additionally, ground-glass opacity has an extensive, contiguous, and amorphous appearance [21]. Our neural network started from the predefined density range but also correctly included areas on the basis of texture (an additional 20% of pixels). These pixels were not detected by the density mask, which was restricted to the predefined density range. The automatic detection of ground-glass opacities required a hierarchic system of multiple networks, operative fields, and filters that have already been successfully applied to other complex problems [13, 22]. Advantages of this system include the consideration of original density information and textural features, self-organization, and hierarchical structure.

On a per-patient and per-scan basis, our system had a sensitivity and negative predictive value of 100% for the detection of ground-glass opacity; for the area involved, the sensitivity and negative predictive value were 99%. These results fulfill the requirements (not to overlook any areas of ground-glass opacity) for a screening preinterpretation tool. For this purpose, the neural network was slightly superior to the density mask with identical thresholds (sensitivity, 89%; negative predictive value, 98%). At the same time, specificity and positive predictive value of the neural network were clearly superior to the density mask, 83% versus 55% and 78% versus 18%, respectively. This improvement is sufficient to make the neural network acceptable for routine application.

Automatic detection of ground-glass opacity has advantages over subjective visual assessment. Neural networks work independently from viewing conditions such as window settings [21], laser printer settings, ambient light, and underlying diseases that may cause heterogeneous lung density. In such cases, it may be difficult to decide whether the hypo- or hyperatenuating areas represent normal parenchyma [23]. Under these conditions, the output of the network can be used as a second opinion to increase specificity and diagnostic accuracy as has been shown for the detection of interstitial opacities [11] and for the differentiation of interstitial lung disease using chest radiography [15]. The visual assessment of ground-glass opacity comprises extent and severity and is highly interpreter-dependent. Conversely, the neural network provides an objective measure for ground-glass opacity. One could argue that the severity of ground-glass opacity is not adequately considered by the neural network. However, the hierarchic structure of multiple networks facilitates and favors the detection of homogeneous areas with high edge contrasts, which correspond to a higher subjective severity rating. Therefore, the severity is represented in the segmented area.

Our study has some limitations. Segmented areas of extensive ground-glass opacity were up to 5% smaller than those identified by the radiologist, findings that represent systemic bias. This bias is caused by the hierarchic structure of the network, which favors a concentration on the core areas of ground-glass opacity together with a marked delineation between ground-glass opacity and denser structures such as vessels or pleura (Fig. 1A,1B,1C). This property will not affect the comparability of results among different patients or during follow-up examinations. The prevalence of ground-glass opacity in the study population (approximately 50%) reflects the diseases that are investigated by high-resolution CT. Because positive predictive value and negative predictive value are dependent on pretest probability, positive predictive value might be underestimated and negative predictive value might be overestimated.

The neural network was not able to overcome typical pitfalls in the detection of ground-glass opacities such as cardiac and respiratory motion artifacts. Partial volumes of bronchovascular bundles, chest wall, diaphragm, and pleura as well as air—tissue interfaces, which have the same density quality as ground-glass opacities, were the most frequent pitfalls. This pitfall has not been described in major publications because it does not pose a real problem to radiologists. Approximately 10% of the ground-glass opacities were attributed to simple dependent opacities. If ground-glass opacities segmented by the neural network had always been regarded as abnormal, subsequent clinical treatment would have been different from that following the evaluation of the radiologist. Although neural networks are sufficiently accurate to be used for the preinterpretation of ground-glass opacities, the final diagnosis of pneumonic infiltrates, alveolitis, fibrosis, dependent opacities, or artifacts is the task of the radiologist.

Supported by the German Research Council (DFG) under grant numbers Th 315/6 and 316/10.

This article contains parts of the doctoral thesis of Dirk Marwede, Johannes Gutenber-University Mainz, 2000.

Address correspondence to H.-U. Kauczor.

We thank Patty Meinhardt for her editorial assistance in preparing this manuscript.

References
Previous sectionNext section
1. Delorme S, Keller-Reichenbecher MA, Zuna I, Schlegel W, van Kaick G. Usual interstitial pneumonia: quantitative assessment of high-resolution computed tomography findings by computer-aided texture-based image analysis. Invest Radiol 1997; 32:566-574 [Google Scholar]
2. Brown MS, McNitt-Gray MF, Mankovich NJ, et al. Method for segmenting chest CT image data using an anatomical model: preliminary results. IEEE Trans Med Imaging 1997; 16:828-839 [Google Scholar]
3. Heitmann K, Rueckert S, Heussel CP, Uthmann T, Thelen M, Kauczor H-U. Fuzzy-neuronale Netze zur automatischen Milzvolumen-Berechnung aus Spiral-CT-Daten. Vergleich mit einer Standard Dichtemaske [abstract in English; text in German]. Fortschr Roentgenstr 2000; 172:139-146 [Google Scholar]
4. Henschke C, Yankelevitz D, Mateescu I, Brettle D, Rainey T, Weingard F. Neural networks for the analysis of small pulmonary nodules. Clin Imaging 1997; 21:390-399 [Google Scholar]
5. Kohonen T. Analysis of a simple self-organizing process. Biol Cybern 1982; 44:135-140 [Google Scholar]
6. Heitmann KR, Kauczor H-U, Mildenberger P, Uthmann T, Perl J, Thelen M. Automatic detection of ground glass opacities on lung HRCT using multiple neural networks. Eur Radiol 1997; 7:1463-1472 [Google Scholar]
7. Gurney JW, Swensen SJ. Solitary pulmonary nodules: determining the likelihood of malignancy with neural network analysis. Radiology 1995; 196:823-829 [Google Scholar]
8. Lin JS, Hasegawa A, Freedman MT, Mun SK. Differentiation between nodules and end-on vessels using a convolution neural network architecture. J Digit Imaging 1995; 8:132-141 [Google Scholar]
9. Wu YC, Doi K, Giger ML, Metz CE, Zhang W. Reduction of false positives in computerized detection of lung nodules in chest radiographs using artificial neural networks, discriminant analysis, and a rule-based scheme. J Digit Imaging 1994; 7:196-207 [Google Scholar]
10. Wu YC, Doi K, Giger ML. Detection of lung nodules in digital chest radiographs using artificial neural networks: a pilot study. J Digit Imaging 1995; 8:88-94 [Google Scholar]
11. Monnier-Cholley L, MacMahon H, Katsuragawa S, Morishita J, Ishida T, Doi K. Computer-aided diagnosis for detection of interstitial opacities on chest radiographs. AJR 1998; 171:1651-1656 [Abstract] [Google Scholar]
12. Scott J, Palmer E. Neural network analysis of ventilation-perfusion lung scans. Radiology 1993; 186:661-664 [Google Scholar]
13. Tourassi GD, Floyd CE, Sostman HD, Coleman RE. Artificial neural network for diagnosis of acute pulmonary embolism: effect of case and observer selection. Radiology 1995; 194:889-893 [Google Scholar]
14. Boone JM. Neural networks at the crossroads. Radiology 1993; 189:357-359 [Google Scholar]
15. Ashizawa K, MacMahon H, Ishida T, et al. Effect of an artificial neural network on radiologists' performance in the differential diagnosis of interstitial lung disease using chest radiographs. AJR 1999; 172:1311-1315 [Abstract] [Google Scholar]
16. Giger ML, Bae KT, MacMahon H. Computerized detection of pulmonary nodules in computed tomography images. Invest Radiol 1994; 29:459-465 [Google Scholar]
17. Gurney JW, Lyddon DM, McKay JA. Determining the likelihood of malignancy in solitary pulmonary nodules with bayesian analysis. II. Application. Radiology 1993; 186:415-422 [Google Scholar]
18. Jain A. Cluster analysis. In: Young T, Fu K, eds. Handbook of pattern recognition and image processing. San Diego: Academic Press, 1986:33-57 [Google Scholar]
19. Döhring W, Linke G. Ein Programmsystem zur quantitativen Auswertung von Computertomogrammen unter Anwendung einer digitalen Maskentechnik zur Isolierung interessierender Organe und Organbereich aus der CT-Wertematrix. Fortschr Röntgenstr 1986; 144:135-148 [Google Scholar]
20. Boone JM, Gross GW, Greco-Hunt V. Neural networks in radiologic diagnosis. I. Introduction and illustration. Invest Radiol 1990; 25:1012-1016 [Google Scholar]
21. Remy-Jardin M, Remy J, Giraud F, Wattinne L, Gosselin B. Computed tomography assessment of ground glass opacity: semiology and significance. J Thorac Imaging 1993; 8:249-264 [Google Scholar]
22. Tsujii O, Freedman MT, Mun SK. Automated segmentation of anatomic regions in chest radiographs using an adaptive-size hybrid neural network. Med Phys 1998; 25:998-1007 [Google Scholar]
23. Im J-G, Kim IH, Chung MJ, Koo JM, Han MC. Lobular low attenuation of the lung parenchyma on CT: evaluation of forty-eight patients. J Comput Assist Tomogr 1996; 20:756-762 [Google Scholar]

Recommended Articles

Automatic Detection and Quantification of Ground-Glass Opacities on High-Resolution CT Using Multiple Neural Networks

Full Access,
American Journal of Roentgenology. 2020;214:1280-1286. 10.2214/AJR.20.22954
Abstract | Full Text | PDF (884 KB) | PDF Plus (895 KB) 
Full Access, , , ,
American Journal of Roentgenology. 2020;214:1072-1077. 10.2214/AJR.20.22976
Abstract | Full Text | PDF (747 KB) | PDF Plus (792 KB) 
Full Access, , ,
American Journal of Roentgenology. 2020;215:87-93. 10.2214/AJR.20.23034
Abstract | Full Text | PDF (705 KB) | PDF Plus (717 KB) | Supplemental Material 
Full Access, , , ,
American Journal of Roentgenology. 2020;214:1078-1082. 10.2214/AJR.20.22969
Abstract | Full Text | PDF (761 KB) | PDF Plus (730 KB) 
Full Access, , ,
American Journal of Roentgenology. 1993;160:249-251. 10.2214/ajr.160.2.8424326
Abstract | PDF (574 KB) | PDF Plus (383 KB) 
Full Access, , ,
American Journal of Roentgenology. 2020;214:1287-1294. 10.2214/AJR.20.22975
Abstract | Full Text | PDF (929 KB) | PDF Plus (911 KB)