 |
 |
|
|
|
Jeffrey E. Arle, M.D., Ph.D., Craig Morriss, M.D., Zhiyue J. Wang, Ph.D., Robert A. Zimmerman, M.D., Peter G. Phillips, M.D., and Leslie N. Sutton, M.D.
Departments of Neurosurgery, Neuroradiology, and Neurooncology, Children's Hospital of Philadelphia; and Department of Neurosurgery, The Hospital of the University of Pennsylvania, Philadelphia, Pennsylvania
Recent studies have explored characteristics of brain tumors by means of magnetic resonance spectroscopy (MRS) to increase diagnostic accuracy and improve understanding of tumor biology. In this study, a computer-based neural network was developed to combine MRS data (ratios of N-acetyl-aspartate, choline, and creatine) with 10 characteristics of tumor tissue obtained from magnetic resonance (MR) studies, as well as tumor size and the patient's age and sex, in hopes of further improving diagnostic accuracy.
Data were obtained in 33 children presenting with posterior fossa tumors. The cases were analyzed by a neuroradiologist, who then predicted the tumor type from among three categories (primitive neuroectodermal tumor, astrocytoma, or ependymoma/other) based only on the data obtained via MR imaging. These predictions were compared with those made by neural networks that had analyzed different combinations of the data. The neuroradiologist correctly predicted the tumor type in 73% of the cases, whereas four neural networks using different datasets as inputs were 58 to 95% correct. The neural network that used only the three spectroscopy ratios had the least predictive ability. With the addition of data including MR imaging characteristics, age, sex, and tumor size, the network's accuracy improved to 72%, consistent with the predictions of the neuroradiologist who was using the same information. Use of only the analog data (leaving out information obtained from MR imaging), resulted in 88% accuracy. A network that used all of the data was able to identify 95% of the tumors correctly. It is concluded that a neural network provided with imaging data, spectroscopic data, and a limited amount of clinical information can predict pediatric posterior fossa tumor type with remarkable accuracy.
Key Words * neural network * pediatric brain tumor * posterior fossa tumor * magnetic resonance spectroscopy * magnetic resonance imaging * children
Since its commercial introduction in 1982, magnetic resonance (MR) imaging has become the neuroradiological study of choice for characterizing brain tumor tissue. Despite tremendous advances in displaying anatomical detail, however, the neuroradiologist's ability to predict tumor histology accurately is still limited when using MR imaging and computerized tomography scanning alone.[7,8,33] Recently, magnetic resonance spectroscopy (MRS) has been explored as a method to characterize further tissue properties in pediatric brain tumors.[26,27,29] Metabolite ratios measured by our group[27,29] using MRS provide information that correlates well with histological findings independent of standard imaging features of pediatric posterior fossa tumors. Although useful characteristics were noted on MRS studies, their accuracy in separating the three predominant histological types (primitive neuroectodermal tumor [PNET], astrocytoma, and ependymoma) by discriminant analysis was no more than 80% when spectra ratios of choline (Cho), creatine (Cr), and N-acetyl-aspartate (NAA) alone were used.[29]
Neural networks are a relatively recent development in the analysis of multivariate problem solving.[12] They are especially well suited to nonlinear pattern recognition problems and have been applied over the past 5 years to a wide range of medical and scientific issues.[1,10,11,16,18,29,3032]
The purpose of the present study was to compare the diagnostic accuracy of readings and predictions regarding pediatric posterior fossa tumors by a trained neuroradiologist (using preoperative MR images), with the predictions of various neural networks designed to train on MR image properties, age of the patient, and spectroscopic data.
CLINICAL MATERIAL AND METHODS
Patient Population
Thirty of the patients had been included in a prior study[29] and three new patients were added. The protocol was approved by the Committee on Human Subjects and informed consent was obtained from parents in all cases. One patient had been excluded from the prior study because the posterior fossa mass was a large teratoma within grossly aberrant newborn anatomy and did not conform to the size and location criteria of that study. Eligibility criteria for the present investigation included the presence of a newly diagnosed cerebellar tumor larger than 10 cm3 that had not been treated previously with surgery or adjuvant therapy, and the ability to undergo MR and MRS studies with or without sedation as necessary. Overall, 33 children ranging in age from 6 months to 14 years were included. All patients underwent surgery within 7 days of the MR studies, and tissue was obtained that allowed pathological correlation with the imaging studies. In general, patients younger than 10 years of age were sedated with pentobarbital (older than 18 months) or chloral hydrate (younger than 18 months) to obtain images that were as free from motion artifact as possible. Spectroscopy typically added approximately 20 to 30 minutes to the overall study time.
Magnetic Resonance Spectroscopy
Studies were performed using a 1.5-tesla MR imager (Magnetom SP; Siemens Corp., Iselin, NJ) equipped with a circular polarized head coil. The spectroscopic data were obtained using either the stimulated-echo acquisition mode or the point-resolved spectroscopy mode, as described in detail in a previous study.[29] The location of the volume of interest was chosen based on the noncontrast MR images obtained just before MRS. None of the spectra were obtained after contrast had been administered.
The three major previously reported spectral components from brain tissue[27,29] were used as analog data in the form of ratios (Cr/NAA, NAA/Cho, Cr/Cho). A description of how the spectroscopy data were further processed, including eddy current correction, apodization, and peak area calculation, can be found in Wang, et al.[29]
Neuroradiological Assessment
The MR images were read by one of the authors (C.M.), who had no previous exposure to the cases and was used in a blinded fashion as a control reader/predictor. Eleven imaging characteristics were noted for each patient in digital form (0 or 1) indicating either presence or absence of a characteristic. These included: midline or mostly hemispheric location; proton density, T1-, and T2-weighted intensities (hypo-, iso-, or hyperintense) relative to normal gray matter; whether cystic or solid; enhancement; edema; flow voids; hemorrhage, focal versus diffuse involvement; and hydrocephalus (Tables 14).
Ultimately, T1-weighted intensity was excluded as a contributing variable because every case exhibited T1-weighted intensity of the mass, which was hypointense relative to the gray matter. It was not possible to know a priori which variables would have useful information in predicting these tumor types. Therefore, we assumed that any variable that was not the same for every case might be valuable, including spectroscopy data that had already been shown to have predictive value in previous studies.[29] This left 10 characteristics that could be used in developing networks.
Analog measurements of the largest axial diameter of the tumors were made and used as a characteristic. The age (to the nearest 6 months) was also noted, as was the sex of the patient. None of the spectroscopy data and none of the histological or gross pathological descriptions of the resected tissue were known by the reading neuroradiologist. After reading and noting each of the criteria, a prediction of whether the tumor was an astrocytoma, PNET, or ependymoma/other was made.
Neural Network Development
Standard back-propagation algorithms and many variants were explored in attempts to develop the most accurate network predictions of tumor types. All implementations of network design and function were performed using commercially available software (NeuralWorks; Neuralware Inc., Pittsburgh, PA). This software allows a wide range of parameters to be adjusted empirically and analyzed to optimize network architectures and learning paradigms. Typically, small networks with one or two hidden layers were analyzed. Results were obtained using various combinations of learning parameters and the network was expanded, both in the number of units within the hidden layer and in the number of hidden layers. The decision level for the output layer nodes was set at 0.9. At each juncture, a variety of learning algorithms were tested systematically. Most often, however, the only adjustment that significantly improved results was a change in the number of nodes in the network. In this fashion, a determination of the "best" network was made for a given set of inputs. All of a network's predictions for a given training set were obtained by cross-validation using a randomly held-aside set of the original data to test a network's predictive abilities. This method is considered to be an effective technique for network validation.[1,2,9] The best networks were cross-validated five times, each time with a different randomly selected test set, and the predictive results were averaged. This was a second layer of validation that helped eliminate the possibility that a given network coincidentally predicted well (or poorly) on a single test set. All test sets were composed of nine cases: three astrocytomas, four PNETs, and two ependymoma/other tumors.
The difficulty with these data was their relative paucity. Specifically, although the data were fairly consistent among tumor groups, the small number of ependymomas (four) and of cases overall with which to train a network (diminished further by the small number representing each group that always had to be set aside for testing) allowed many networks to learn the training sets perfectly but left them unable to predict the test cases well. This was often partly caused by overtraining of a given network[2,3] but was difficult to avoid even when networks were appropriately trained.
To deal with this problem of sample size, additional "sample" cases were created by using the original input data and adding a small amount of Gaussian noise. In this way, 50 examples of each tumor type were produced and used to train the networks. This technique has been used successfully by others[15,22] in the context of using neural networks with small sample sizes. Testing the networks for predictive ability was only conducted on actual case data, however. Attempts were made to segregate the tumors into four groups by separating ependymomas into an individual group. This could not be accomplished with any reliability, however, because the number of real examples for these groups was prohibitively small.
Several attempts were made to train networks using only the analog data (that is, age, tumor size, and the spectra ratios). This was done to determine whether imaging characteristics carried any weight in helping to determine the tumor type. Attempts were also made to find network architectures and learning parameters that predicted tumor types successfully using only the image criteria and ignoring analog data. It should be noted that most of the networks that were attempted, and all networks analyzed in depth learned training sets with 100% accuracy. The results shown pertain only to a network's ability to predict the nine cases that were purposely omitted from the training set. (See Appendix for the mathematical basis underlying neural network learning, back-propagation, and the variations used in this study.)
RESULTS
The initial step in evaluating the MR imaging data involved obtaining readings from a trained neuroradiologist who was unfamiliar with the cases. This was done in a blinded manner in which the neuroradiologist was unaware of the patient's name or diagnosis and gave interpretations for each imaging variable (Tables 13). A predicted single diagnosis was then given as well. Table 5 shows the results of this evaluation.
Table 6 shows further statistical measures of the accuracy of the neuroradiologist's readings. These data provide a more direct comparison with the effectiveness of other studies in which different methods of diagnosis may have been used. The measures noted here include specificity and sensitivity, as well as positive and negative predictive values. The results reveal the nature of errors made in MR image interpretation. The PNETs and ependymomas were largely mistaken for each other when errors were made in their respective categories, whereas astrocytomas had a positive predictive value of 100%. However, when missed, astrocytomas could be mistaken for being either PNETs or ependymomas. The lowest positive predictive value occurred with ependymomas (0.57), although specificity was highest with this group (0.81).
Four networks were developed by means of systematic manipulation of network parameters, with each group of data being evaluated. A partial comparison of the network structures is given in Table 7. The number of interconnection weights is given rather than a listing of every weight in a specific network.
The most robust confirmation of neural network predictive abilities involves cross-validation, in which the network is trained on a portion of the data and tested on real data that is not included in the training set. In addition, multiple cross-validations may be used to show internal consistency for a network. This was done in all cases in these studies and results are shown in Table 8. Cross-validation was performed five times in each network, using a different subset of the data each time for training and testing. The mean percentages and standard deviations of the networks' predictive accuracy are also shown for each network.
DISCUSSION
The result that the major tumor types in the posterior fossa in children could be well differentiated in a neural network by simple MR imaging, MRS, and limited clinical data is surprising with so few cases on which to train. This potential difficulty in using neural networks for medical diagnosis has been noted by others as well.[2,3,9,13,14,19,31,32] Astion and Wilding[3] analyzed much of the medical diagnostic literature that used neural networks published prior to their study and found that 1) sample size is universally problematic; 2) there are only a few insights into optimizing a network or even estimating its appropriate size, and 3) there were few examples in which the neural network analysis had high predictive abilities (gt 85%). Wu, et al.,[31] found better than 90% accuracy in classifying microcalcification clustering in mammograms using 56 cases each in three categories with networks requiring 32 to 1024 inputs (for graphic data) and five to 15 hidden nodes. Their analysis used both cross-validation and receiver operating characteristic analysis. On the other hand, O'Leary, et al.,[19] used a sparse three-layer network to analyze only 36 cases of breast cancer (either tubular carcinoma or sclerosing adenoma) and found 100% accurate test case prediction, although it is unclear whether they had used samples from the training set in the test set and whether they cross-validated results more than once, because their network required a tremendous number of iterations (460,000) just to train, indicating that it may have been overtrained.
Following the rough guidelines of Masters[14] and those suggested by others,[17] we should have required approximately 1200 cases to predict these tumors reliably with 20 inputs and two hidden layers having 20 and seven nodes, respectively (Table 7). Masters suggests taking the number of weights in the network and doubling it to estimate the number of training cases needed. In any event, despite having essentially 150 training cases (once the artificial cases were created by adding noise), the original small number of actual cases may not have broadened our example pool enough to obtain 100% accuracy consistently. We did find that when more artificial cases were created and used, the results did not vary significantly.
Our results indicate that as more data are used, the predictive ability of the network improves, especially if the data are analog. When using only the three spectroscopy ratios, a network could predict almost 60% of the cases. Adding only patient age and sex and tumor size to the spectroscopy data allowed a network to achieve almost as much predictive ability as using all of the data (87.8% vs. 94.6%). This difference in the weight of influence between the analog inputs and the digital inputs is more clearly seen when comparing the results of using all of the data except spectroscopy (which is fundamentally the same as using those data that were available to the neuroradiologist) with the results using analog data alone (71.7% vs. 87.8% predictive accuracy). If we accept that patient age and sex and tumor size taken alone add a relatively fixed percentage to a network's predictive abilities, then it appears that MR imaging data alone account for approximately 40% of predictions. If we reject this speculation, however, and believe that this multivariate system is nonlinearly influenced by age, sex, and lesion size, then these data indicate that MR imaging data may actually confuse readings that would otherwise have been more accurate using just spectra, age, sex, and tumor size.
Efforts were made to find the optimum set of network parameters (learning algorithm, learning coefficients, momentum terms, and numbers of nodes) for a given set of inputs. Although this task involves more art than science, the variables that seemed most relevant were adjusted in a systematic way to find the best results. As mentioned, the results were judged by multiple cross-validations, and it was possible to train almost any network to learn all of the training set data. However, this approach leaves open the possibility that part of a network's predictive success is based on the researcher's ability to create an optimized network. For example, if we had manipulated each variable in network design of the spectroscopy-only network in exactly the appropriate manner, it is possible that we could have obtained a network with 100% predictive ability. Our experience working with many varieties of learning algorithms and manipulations of the variables in creating these neural networks indicates that this is unlikely. We do not know, however, because there is no heuristic for network optimization and little information has been published on ways to test for optimization in neural network design.[4]
As more tumor data accumulates it is likely that at least two enhancements to our analysis will occur. First, the number of input variables needed will decrease. We have noted already that the relative tumor intensity on T1-weighted images played no role in discrimination. Combinations of more sophisticated MR imaging qualities might reasonably be exchanged for the simpler ones used here, thereby reducing the number of inputs needed. However, the danger always remains that adding subtlety to the MR imaging inputs narrows their reliability in widespread use. One interesting possibility is that MRS data alone will be found adequate to differentiate tumor types within a network. In this study we achieved only 58.5% accuracy with spectroscopy alone (three inputs), but with more training sets a network might require only a few input nodes instead of the 20 that we used for the best predictions. Again, adding just age and tumor size increased the predictive ability to a respectable 87.8%. Second, the number of hidden nodes might decrease. As more data become available the number of nodes will converge toward an optimum amount, which is likely to be a lower number of nodes than used here because the outcome space is relatively small and similar problems (in terms of variable number and outcome number) typically require smaller networks when the number of cases in the training sets is adequate.
This study presupposes the value of an accurate preoperative histological diagnosis in pediatric posterior fossa tumors, which will be surgically treated in any event. The surgical management of these tumors is determined by histological findings, because astrocytomas and ependymomas require total excision for cure,[23,24] whereas PNETs do not.[25] Currently, the decision whether to remove tumors from the brainstem is made on the basis of findings on intraoperative frozen sections, which may not be 100% accurate.[5,6,20,21] A preoperative neural network with complete diagnostic accuracy might replace intraoperative histological examination in guiding surgery and have the added advantage of allowing appropriate counseling of parents regarding surgical risk. Another advantage of knowing the tumor's histological type from diagnostic studies is that consideration may be given to preoperative adjuvant chemotherapy, as is done for other tumors. Furthermore, costly and time-consuming spinal staging MR studies for drop metastases may be avoided for patients with typical cerebellar astrocytomas, and they may be obtained preoperatively in cases of ependymoma and PNET, when the tests are most accurate. Finally, the methodology described here may have a role in the diagnosis of tumors in other locations in which accurate noninvasive diagnostic studies could obviate the need for biopsy entirely, such as glioblastoma multiforme, germinoma, or certain brainstem and midline gliomas for which treatment may be primarily medical.
Appendix
The standard neural network learning method typically falls under the term "back-propagation." These networks are used for problems with multiple variables that each contribute an unknown amount to the outcome of the process in question.[12,17] They do not require explicit rules or knowledge external to the problem but rather generate their own rules in a sense. The structure of the network has an input layer made of "nodes" that take on the values of the variables that are known in the problem. They are connected to more nodes that make up what are called "hidden" layers and each connection is associated with a strength or "weight." Usually, only one or two hidden layers are used, and there have been some studies indicating that arbitrarily complex problems may be solved with three or fewer layers. Finally, these hidden layers are connected, again with weighted outputs, to an output layer as shown in the schematic below.
Sometimes the input layer is directly connected to the output layer as well, and sometimes the hidden layer(s) feed back on the input layer as desired or determined by the nature of the problem. Each node then processes its inputs via an internal function to get its own output, as shown in the second schematic, in which Oi(t) is the output of the ith node in layer t and whi(t) is the weight of the connection from the hth node of the prior layer on the ith node of layer t.
Summating these outputs with their weights and performing a so-called "transfer" function on the summation results in the output of the ith node as in the following formula:
Oi(t) = F{sigmai (whi[t] * Oi[t1])} = F (Si[t]),
where F is typically a sigmoidal or hyperbolic tangent function and Si(t) represents the weighted summation of inputs to the ith node in layer t. If the weights are initially randomly assigned throughout the network, the output layer will have an error when compared to the desired output. The key to allowing the network to learn the correct results and to generalize as well to inputs not seen before is to use this error to adjust the weights in such a way as to produce a corrected output. This is accomplished by finding the local error at each node (ei). If one assumes that the overall network has a global error (E) that is a differentiable function of all of the connection weights in the network, the problem is somewhat simplified. This is convenient for functions like the sigmoidal transfer function:
F(x) = (1 + ex)1
where the derivative is the simpler
F'(x) = F(x) * (1F[x])
Because the relationship between E and the sum of the network nodes' outputs is expressed via the partial derivatives as follows:
eit = E/Si(t)
combining equations above yields:
ei(t) = Oi(t) * (1Oi[t]) * sigmaj (ej[t + 1] * Wij[t + 1]).
The ei can then be used to compute the change in each individual weight (w):
whi(t) = C * ei(t) * Oh(t 1)
where C is a coefficient of learning partly determining the speed of learning convergence. These ws are added to each appropriate weight before presentation of the input values again. By this general means, the network will converge in its connections to a set of weights that will produce correct results for the known input variables, and will reasonably generalize for previously unknown input patterns in the same problem space.
References
1. Astion ML, Wener MH, Thomas RG, et al: Application of neural networks to the classification of giant cell arteritis. Arth Rheum 37:760770, 1994
2. Astion ML, Wener MH, Thomas RG, et al: Overtraining in neural networks that interpret clinical data. Clin Chem 39:19982004, 1993
3. Astion ML, Wilding P: The application of backpropagation neural networks to problems in pathology and laboratory medicine. Arch Pathol Lab Med 116:9951001, 1994
4. Baum EB, Haussler D: What size net gives valid generalization? in Touretzky DS (ed): Advances in Neural Information Processing Systems I. San Mateo, CA: Morgan Kaufmann, 1989, pp 8190
5. Cappabianca P, Spaziante R, Caputi F, et al: Accuracy of the analysis of multiple small fragments of glial tumors obtained by stereotactic biopsy. Acta Cytol 35:505511, 1991
6. Cappabianca P, Spaziante R, de Devitiis E, et al: Computerized tomography-guided biopsy. J Neurosurg 82:149150, 1995 (Letter)
7. Cavouras D, Prassopoulos P: Computer image analysis of brain CT images for discriminating hypodense cerebral lesions in children. Med Inform 19:1320, 1994
8. Choksey MS, Valentine A, Shawdon H, et al: Computed tomography in the diagnosis of malignant brain tumors: do all patients require biopsy? J Neurol Neurosurg Psychiatry 52:821825, 1989
9. Ciccheti DV: Neural networks and diagnosis in the clinical laboratory: state of the art. Clin Chem 38:910, 1992
10. Erler BS, Hsu L, Truong HM, et al: Image analysis and diagnostic classification of hepatocellular carcinoma using neural networks and multivariate discriminant functions. Lab Invest 71:446451, 1994
11. Hamamoto I, Okada S, Hashimoto T, et al: Prediction of the early prognosis of the hepatectomized patient with hepatocellular carcinoma with a neural network. Comp Biol Med 25:4959, 1995
12. Lapedes A, Farber R: How Neural Nets Work. Preprint LA-UR-88-418. Los Alamos, New Mexico: Los Alamos National Laboratory, 1988
13. Maclin PS, Dempsey J: How to improve a neural network for early detection of hepatic cancer. Cancer Lett 77:95101, 1994
14. Masters T: Practical Neural Network Recipes in C++. San Diego: Academic Press, 1993
15. Meyer B, Hansen T, Nute D, et al: Identification of the 1H-NMR spectra of complex oligosaccharides with artificial neural networks. Science 251:542544, 1991
16. Moul JW, Snow PB, Fernandez EB, et al: Neural network analysis of quantitative histological factors to predict pathological stage in clinical Stage I nonseminomatous testicular cancer. J Urol 153:16741677, 1995
17. NeuralWare: Neural Computing. A Technology Handbook for Professional II/PLUS and NeuralWorks Explorer. Pittsburgh: Neuralware, Inc., 1995
18. Oghalai JS, Street WN, Rhode WS: A neural network-based spike discriminator. J Neurosci Methods 54:922, 1994
19. O'Leary TJ, Mikel UV, Becker RL: Computer-assisted image interpretation: use of a neural network to differentiate tubular carcinoma from sclerosing adenosis. Mod Pathol 5:402405, 1992
20. O'Neill KS, Dyer PV, Bell BA, et al: Is perioperative smear cytology necessary for CT-guided stereotactic biopsy? Br J Neurosurg 6:421427, 1992
21. Ranjan A, Rajshekhar V, Joseph T, et al: Nondiagnostic CT-guided stereotactic biopsies in a series of 407 cases: influence of CT morphology and operator experience. J Neurosurg 79:839844, 1993
22. Sietsma J, Dow RJF: Creating artificial neural networks that generalize. Neural Networks 4:6779, 1991
23. Sutton LN, Cnaan A, Klatt L, et al: Postoperative surveillance imaging in children with cerebellar astrocytomas. J Neurosurg 84721725, 1996
24. Sutton LN, Goldwein J, Penlongo G, et al: Prognostic factors in childhood ependymomas. Pediatr Neurosurg 16:5765, 19901991
25. Sutton LN, Phillips PC, Molloy PT: Surgical management of medulloblastoma. J Neurooncol 29:921, 1996
26. Sutton LN, Wang Z, Duhaime AC, et al: Tissue lactate in pediatric head trauma: a clinical study using 1H-NMR spectroscopy. Pediatr Neurosurg 22:8187, 1995
27. Sutton LN, Wehrli SL, Gennarelli L, et al: High-resolution 1H-magnetic resonance spectroscopy of pediatric posterior fossa tumors in vitro. J Neurosurg 81:443448, 1994
28. Vertosick FT, Rehn T: Predicting behavior of an enzyme-linked immunoassay model by using commercially available neural network software. Clin Chem 39:24782482, 1993
29. Wang Z, Sutton LN, Cnaan A, et al: Proton MR spectroscopy of pediatric cerebellar tumors. AJNR 16:18211833, 1995
30. Wu C, Shivakumar S: Back-propagation and counter-propagation neural networks for phylogenetic classification of ribosomal RNA sequences. Nucleic Acids Res 22:42914299, 1994
31. Wu Y, Doi K, Giger ML, et al: Computerized detection of clustered microcalcifications in digital mammograms: applications of artificial neural networks. Med Physics 19:555560, 1992
32. Zhang W, Doi K, Giger ML, et al: Computerized detection of clustered microcalcifications in digital mammograms using a shift-invariant artificial neural network. Med Physics 21:517524, 1994
33. Zimmerman RA, Bilaniuk LT, Rebsamen S: Magnetic resonance imaging of pediatric posterior fossa tumors. Pediatr Neurosurg 18:5864, 1992
Manuscript received May 22, 1996.
Accepted in final form November 26, 1996.
Address reprint requests to: Leslie N. Sutton, M.D., Department of Neurosurgery, Children's Hospital of Philadelphia, 34th Street and Civic Center Boulevard, Philadelphia, Pennsylvania 19104.
Click here to return to Journal of Neurosurgery Table of Contents.
Click here to return to Neurosurgical Focus Table of Contents.
|
|
 |