A Generic Support Vector Machine Model for Preoperative Glioma Survival Associations
Abstract
Purpose
To develop a generic support vector machine (SVM) model by using magnetic resonance (MR) imaging–based blood volume distribution data for preoperative glioma survival associations and to prospectively evaluate the diagnostic effectiveness of this model in autonomous patient data.
Materials and Methods
Institutional and regional medical ethics committees approved the study, and all patients signed a consent form. Two hundred thirty-five preoperative adult patients from two institutions with a subsequent histologically confirmed diagnosis of glioma after surgery were included retrospectively. An SVM learning technique was applied to MR imaging–based whole-tumor relative cerebral blood volume (rCBV) histograms. SVM models with the highest diagnostic accuracy for 6-month and 1-, 2-, and 3-year survival associations were trained on 101 patients from the first institution. With Cox survival analysis, the diagnostic effectiveness of the SVM models was tested on independent data from 134 patients at the second institution. Results were adjusted for known survival predictors, including patient age, tumor size, neurologic status, and postsurgery treatment, and were compared with survival associations from an expert reader.
Results
Compared with total qualitative assessment by an expert reader, the whole-tumor rCBV-based SVM model was the strongest parameter associated with 6-month and 1-, 2-, and 3-year survival in the independent patient data (area under the receiver operating characteristic curve, 0.794–0.851; hazard ratio, 5.4–21.2).
Discussion
Machine learning by means of SVM in combination with whole-tumor rCBV histogram analysis can be used to identify early patient survival in aggressive gliomas. The SVM model returned higher diagnostic accuracy values than an expert reader, and the model appears to be insensitive to patient, observer, and institutional variations.
© RSNA, 2014
Introduction
To answer the call for improved reproducibility and standardization in clinical imaging (1), as well as the increasing complexity of cancer imaging protocols, machine learning methods have been introduced as a computer-aided diagnostic tool in glioma characterization for reduced operator measurement error (2–5). According to the assumption that high malignancy is reflected by increased vascular growth and tortuosity (6,7), glioma grading or survival associations can be assessed by using the quotient of relative cerebral blood volume (rCBV) values in tumor hot spots and reference tissue or whole-tumor rCBV distribution analysis (8–12). Reported diagnostic accuracies in single-institution and retrospective data vary from 70% to 90% and warrant further research on improved repeatability (2,3,8,9). When compared with most traditional radiographic measures of tumor malignancy (contrast material enhancement and mean values of diffusion and perfusion magnetic resonance [MR] imaging), whole-tumor rCBV analysis in combination with support vector machines (SVMs) have previously been shown to be the favorable univariate approach for preoperative survival associations in adult patients with gliomas (5). Both tumor grading (4) and survival predictions (5) from rCBV-based SVM analysis of glioma patient data show accuracy values in the range of 75%–85%. However, undeterred by sophisticated correction algorithms for overfitting and multiple comparisons, the value of a general SVM model for use in independent patient glioma data from a different institution has not been assessed.
To this end, the aim of our study was to develop a generic rCBV-based SVM model for preoperative glioma survival associations and prospectively evaluate the diagnostic effectiveness of this model in autonomous patient data.
Materials and Methods
One author (A.B.) is a board member of NordicNeuroLab, Bergen, Norway. Authors without a financial interest controlled the data and information that could have caused a conflict of interest.
Patients
Our retrospective study was approved by institutional and regional medical ethics committees, and all patients signed a consent form. From May 2003 to July 2012, our binational study included 101 untreated patients from institution A (48 women, 53 men; median age, 53 years; range, 18–86 years) and 134 untreated patients from institution B (61 women, 73 men; median age, 55 years; range, 18–87 years) who were referred for standard MR imaging examination, including contrast-enhanced perfusion imaging (Fig E1 [online]). At both institutions, this MR imaging protocol is part of the diagnostic decision-making process for all incoming patients suspected of having an intra-axial brain lesion. An overview of histopathologic diagnoses, Karnofsky performance status, and treatments are shown in Table E1 (online). The 235 patients have not been included in any previous study on perfusion MR imaging or SVM analysis, and all patient identifiers were removed prior to study inclusion. Survival outcomes were reviewed up until August 2013 by using in-house and national registry systems, as well as patient medical records. Surviving patients and those who died of non–glioma-related diseases were defined as censored cases. Censoring was performed for institution A only (21 patients alive at <2 years and 29 patients alive at <3 years). No censoring was performed on the institution B data, which were complete at 3 years.
MR Imaging and Postprocessing
Imaging was performed with 1.5-T and 3-T imaging units at both institutions, with eight-, 12-, or 32-channel head coils used for Siemens MR imaging units and eight- or 16-channel head coils used with GE Medical Systems MR imaging units (Table E2 [online] contains a full list of MR imaging units, manufacturers, and imaging parameters used). Tumor outlines were drawn semiautomatically on conventional MR images by a board-certified neuroradiologist with 10 years of clinical experience (M.C.P., hereafter referred to as the “expert reader”), who was blinded to relevant patient data and used the open source platform “3D Slicer” (13). Enhancing lesions on axial contrast-enhanced spin-echo T1-weighted images (repetition time msec/echo time msec, 500/8) and hyperintensities on fluid-attenuated inversion-recovery images (9000/110; inversion time, 1500 msec) and fast spin-echo T2-weighted images (4000/100) were included, while taking care to exclude regions of necrosis and cysts (9,14) (Fig 1a). The neuroradiologist spent a mean of 7 minutes outlining tumors for each patient. Nontumor macroscopic vessels were excluded by using an automatic segmentation algorithm on dynamic susceptibility contrast-enhanced MR images (15). By using established procedures (14), rCBV maps were created automatically in nordicICE (NordicNeuroLab, Oslo, Norway) without human interaction according to the area under the contrast-enhanced first-pass data, normalized to reference tissue and corrected for contrast agent extravasation (5,9,14). Whole-tumor rCBV histograms were created by using 100 bins over an rCBV range of 0–5 arbitrary units (9) (Fig 1b). To reduce the effect of outliers, all rCBV values outside the 95th percentile were excluded. The mean number of days from MR imaging to surgery ± standard error was 6 days ± 1 for institution A and 9 days ± 3 for institution B.
Figure 1a: Images demonstrate rCBV histogram analysis in gliomas. (a) On the anatomic T2-weighted MR image and rCBV map overlay for a 26-year old woman with a grade II oligodendroglioma, the outline of the tumor region used for rCBV histogram analysis is shown in red. (b) The resultant whole-tumor rCBV histogram from a is given with the average rCBV histograms (±1.96 standard errors [SE]) of all grade II gliomas and grade IV gliomas from institution A, respectively. In contrast to the relatively homogeneous rCBV distributions of the low-grade gliomas, the average rCBV histogram of the grade IV gliomas depicts a wide distribution attributed to the increased and heterogeneous vascularity of these tumors.
Figure 1b: Images demonstrate rCBV histogram analysis in gliomas. (a) On the anatomic T2-weighted MR image and rCBV map overlay for a 26-year old woman with a grade II oligodendroglioma, the outline of the tumor region used for rCBV histogram analysis is shown in red. (b) The resultant whole-tumor rCBV histogram from a is given with the average rCBV histograms (±1.96 standard errors [SE]) of all grade II gliomas and grade IV gliomas from institution A, respectively. In contrast to the relatively homogeneous rCBV distributions of the low-grade gliomas, the average rCBV histogram of the grade IV gliomas depicts a wide distribution attributed to the increased and heterogeneous vascularity of these tumors.
The expert reader provided a qualitative prediction of survival (alive or deceased) for 6 months and 1, 2, and 3 years on the basis of clinical (age and Karnofsky performance status) and structural MR imaging data (including apparent diffusion coefficient maps from diffusion MR imaging) (11,12,16), taking into account his own experience and known factors of prognostic significance for gliomas (17–20). To best mirror a traditional radiographic assessment, the reader was blinded to perfusion data, histopathologic findings, and treatment information.
SVM Analysis
We used a previously described algorithm for survival predictions that has been found particularly useful for histogram-based data—the ν-SVM algorithm (5,21,22). The ν-SVM algorithm (Appendix E1 [online]) was implemented in Matlab 2013a (MathWorks, Natick, Mass), including the latest version of the “libsvm” toolbox (21). The ν-SVM model is parameterized by inputs ν and λ (21). The ν parameter determines the softness of the classification margins and is regarded as a penalty parameter for misclassifications. The optimal ν value is found by means of iterative sampling over the parameter range (0–1), resulting in a tradeoff between model generalization and predictive accuracy (21). The λ parameter regulates the kernel width of a radial basis function and is fixed to 1/size of the feature space (ie, number of rCBV histogram bins) (5). The SVM model was trained by using the whole-tumor rCBV histogram data sets of the 101 subjects from institution A, and a 10-fold cross-validation procedure was performed to avoid overfitting. The SVM model was adjusted for patient ages and tumor sizes estimated from the MR imaging outlines by including these two covariates as additional input features. Finally, the procedure described earlier was performed four times to create four separate SVM models with clinically relevant survival ranges of 6 months and 1, 2, and 3 years. For any new patient, the SVM models will thus return a prediction representative of the most probable outcome at the different survival ranges.
To evaluate the SVM procedure with independent data, the models were applied to the whole-tumor rCBV histograms of the 134 subjects from institution B, with patient age and tumor volume given as covariate features. None of these data were available during training of the SVM models. Also, to further assess the robustness of the SVM procedure, it was reversed, and data from institution B were used to predict survival at institution A.
Statistical Analysis
The diagnostic accuracy of 6-month and 1-, 2-, and 3-year preoperative SVM-based survival associations (alive or deceased) was assessed by using values of sensitivity, specificity, and area under the receiver operating characteristic curve (AUC), as well as Kaplan-Meier analysis and hazard ratios from Cox regression. Results from the SVM analysis were compared with the survival associations predicted by the expert reader by using the McNemar test. The significance level was P less than .05, with Holm-Bonferroni corrections for multiple comparisons (four stepwise adjustments). To adjust for varying and institution-specific treatment routines (19), covariates of the Cox regression included patient sex, Karnofsky performance status (percentage), and steroid use at the time of MR imaging (“yes” or “no”), as well as time-dependent covariates for the time and type of surgery (biopsy, subtotal [<90%] resection, or gross total [>90%] resection), the time and type of adjuvant therapy (radiation therapy, chemotherapy, or combined chemotherapy and radiation therapy with or without concomitant antiangiogenic therapy), and, finally, time and presence of occurrence (“yes” or “no”) and/or repeated surgery. Statistical analysis was performed by using SPSS version 18 software (SPSS, Chicago, Ill).
Results
Patient survival outcomes and corresponding SVM-based diagnostic accuracy values and model parameters for 6-month and 1-, 2-, and 3-year survival associations for patients from institution B are shown in the Table.
 |
When compared with the expert reader (Table E3 [online]), the whole-tumor rCBV-based SVM model was the strongest parameter associated with 6-month and 1-, 2-, and 3-year survival in the independent patient data (McNemar test, P < .01 for all time points). Illustrations of the SVM model hyperplanes and resulting Kaplan-Meyer survival curves for the shortest (6-month) and longest (3-year) survival associations for SVM and the expert reader are shown in Figure 2a and 2b and Figure 2c and 2d, respectively. The corresponding Kaplan-Meyer survival curves for 1- and 2-year survival associations are shown in Figure E2 (online).
Figure 2a: Plots and graphs represent support vectors for glioma survival predictions. (a) Plot shows a two-dimensional projection of the high-dimension SVM space used for 6-month survival analysis. The higher-order support vectors ( X ) and hyperplane (solid line) are projected in two dimensions by using principal component analysis. The resulting hyperplane is the optimal tradeoff between high model generalization and prediction accuracy. Red dots represent patients alive at 6 months, and blue dots represent deceased patients. (b) Corresponding Kaplan-Meier 6-month survival curves were derived from the rCBV-based SVM model, as well as from the expert reader for reference. (c) Plot shows the corresponding two-dimensional principal component analysis of the high-dimension SVM space used for 3-year survival analysis. (d) Corresponding Kaplan-Meier 3-year survival curves were derived from the rCBV-based SVM model and expert reader. NA = not applicable.
Figure 2b: Plots and graphs represent support vectors for glioma survival predictions. (a) Plot shows a two-dimensional projection of the high-dimension SVM space used for 6-month survival analysis. The higher-order support vectors ( X ) and hyperplane (solid line) are projected in two dimensions by using principal component analysis. The resulting hyperplane is the optimal tradeoff between high model generalization and prediction accuracy. Red dots represent patients alive at 6 months, and blue dots represent deceased patients. (b) Corresponding Kaplan-Meier 6-month survival curves were derived from the rCBV-based SVM model, as well as from the expert reader for reference. (c) Plot shows the corresponding two-dimensional principal component analysis of the high-dimension SVM space used for 3-year survival analysis. (d) Corresponding Kaplan-Meier 3-year survival curves were derived from the rCBV-based SVM model and expert reader. NA = not applicable.
Figure 2c: Plots and graphs represent support vectors for glioma survival predictions. (a) Plot shows a two-dimensional projection of the high-dimension SVM space used for 6-month survival analysis. The higher-order support vectors ( X ) and hyperplane (solid line) are projected in two dimensions by using principal component analysis. The resulting hyperplane is the optimal tradeoff between high model generalization and prediction accuracy. Red dots represent patients alive at 6 months, and blue dots represent deceased patients. (b) Corresponding Kaplan-Meier 6-month survival curves were derived from the rCBV-based SVM model, as well as from the expert reader for reference. (c) Plot shows the corresponding two-dimensional principal component analysis of the high-dimension SVM space used for 3-year survival analysis. (d) Corresponding Kaplan-Meier 3-year survival curves were derived from the rCBV-based SVM model and expert reader. NA = not applicable.
Figure 2d: Plots and graphs represent support vectors for glioma survival predictions. (a) Plot shows a two-dimensional projection of the high-dimension SVM space used for 6-month survival analysis. The higher-order support vectors ( X ) and hyperplane (solid line) are projected in two dimensions by using principal component analysis. The resulting hyperplane is the optimal tradeoff between high model generalization and prediction accuracy. Red dots represent patients alive at 6 months, and blue dots represent deceased patients. (b) Corresponding Kaplan-Meier 6-month survival curves were derived from the rCBV-based SVM model, as well as from the expert reader for reference. (c) Plot shows the corresponding two-dimensional principal component analysis of the high-dimension SVM space used for 3-year survival analysis. (d) Corresponding Kaplan-Meier 3-year survival curves were derived from the rCBV-based SVM model and expert reader. NA = not applicable.
When reversing the SVM analysis, the SVM model from institution B returned similar diagnostic accuracy values for the 6-month and 1-, 2-, and 3-year survival associations of institute A (hazard ratios of 4.6, 10.1, 1.9, and 2.75, respectively; and AUCs of 0.775, 0.702, 0.703, and 0.803, respectively).
Discussion
Our results show that the whole-tumor rCBV-based SVM model returned higher diagnostic accuracy values than an expert reader for 6-month to 3-year survival associations in a large cohort of independent data from a collaborating institution on a different continent, both institutions with unique diagnostic and treatment protocols. Because most imaging biomarkers for diagnostic accuracy and response in cancer are created and validated on the basis of data from the same cohort or institution (11,12), our findings therefore help address the inherent study bias associated with data dependency from imaging hardware, software, image analysis algorithms, and patient selection criteria, as well as national and local guidelines for clinical use and postsurgical treatment strategies (17,23). By evaluating our model in a carefully planned and large multicenter study (16), the true value of our imaging biomarker is demonstrated.
Our study builds on previous efforts to define pattern recognition models for MR imaging characterization of patients with glioma (2–5,24,25) and leaves us with two key advances in knowledge. First, the hemodynamic features of highly aggressive and malignant gliomas that lead to patient death within 6 months are distinctly different from those in patients with less malignant tumors and overall survival beyond 6 months. Most of the cancers in the short-survival group were glioblastomas (but not exclusively), which are known to exhibit a wide range of histopathologic and mitotic genetic signatures that influence both imaging and survival (18). This result is in line with other work in which the predictive value of apparent diffusion coefficient histograms (12,16), conventional contrast enhancement (17), necrosis at MR spectroscopy (12), and regions of maximum rCBV (6,8,12) were quantified. Interestingly, the ν parameter in the SVM model increased from 0.425 (at 6 months) to 0.675 (at 2 years) before returning to 0.4 at 3 years. Larger values of ν allow more data points to lie within the classification margins, which in practice, corresponds to a more general SVM model at the potential cost of more misclassifications (22). The relatively low ν value of 0.4 at 6 months demonstrates tight classification margins with few options to expand—that is, the remaining misclassified patients have characteristics that are sparsely distributed in the surrounding feature hyperspace. At 1- and 2-year survival, the ν value increased without a leap in diagnostic accuracy. This suggests that the histopathologic and genetic profiles and, thus, the overall survival probabilities at these intervals are more evenly distributed. Similarly, the reversal of the 3-year ν value indicates tighter SVM model margins, which is consistent with the 10% survival rate of glioblastomas at 3 years (19).
The second key finding is that the presented SVM approach returned diagnostic accuracy values comparable to those of the single-institution study by Emblem et al (5) (1–4-year AUC, 0.77–0.82) and therefore appears relatively insensitive to the specific variations in patient demographics, glioma type, and imaging acquisition routines, as well as differences in pre- and postoperative treatment strategies. For example, the number of patients with low-grade gliomas relative to high grades were approximately twice as high in institution A as in institution B. Reasons for this include time-dependent and stochastic variations in patient admittance, skewness in regional and national patient coverage, and, probably most important, institution profile and local expertise (17). An attractive benefit of this finding is that institutions without knowledge or resources to create a new SVM model can probably adapt the current SVM model to their patient data by using the proposed model parameters. This hypothesis, however, needs to be confirmed in independent and prospective studies.
Our study may be limited by the preselection of only patients with glioma. Untreated differential diagnoses should be included in a prospective study to identify unique clinical and imaging features of brain cancer subtypes (2,4). However, for a histologically confirmed diagnosis of glioma, our work suggests that the preferred biomarker for survival associations is actually to go back to the preoperative perfusion MR imaging examination and perform additional SVM analysis. In this setting, the added value of postsurgical imaging and clinical information should also be evaluated in future studies. Furthermore, because most institutions have their own preferred methods and guidelines for tumor definitions, we used the traditional method of semiautomatic outlining by an experienced neuroradiologist. In addition to the minimal observer variability associated with a whole-tumor histogram-based approach (9), we believe semiautomatic outlining is a feasible and generally acceptable approach for identification of tumors in this setting, especially at nonacademic institutions where there is no experience with advanced computer-aided tumor segmentation. Therefore, our findings should be suitable for most institutions. We propose that the SVM approach used in our study may be successfully complemented by automatic tumor segmentation for improved characterization of patients with glioma (3,5,10).
In conclusion, we present a generic SVM model for preoperative glioma survival associations. The suggested list of ν-SVM parameters is a starting point to develop similar machine learning models. The described SVM model, combined with perfusion-weighted MR imaging, is an attractive alternative to traditional methods for upfront assessment of survival outcome and treatment planning in adult patients suspected of having or confirmed to have glioma.
Advances in Knowledge
■ Support vector machines, in combination with MR imaging–based blood volume analysis, show that an advanced machine learning technique can be used to identify 6-month to 3-year patient survival (area under the receiver operating characteristic curve [AUC], 0.794 to 0.851, respectively) in aggressive gliomas with higher accuracy than an expert reader (AUC, 0.496 to 0.658, respectively).
■ The machine learning technique presented in our study appears to be insensitive to the institutional variations in imaging acquisition routines and treatment strategies.
Implication for Patient Care
■ Machine learning techniques have the potential to enhance the availability and standardization of current advanced MR imaging methods for preoperative glioma characterization and to therefore improve treatment planning.
Acknowledgment
The authors thank David Scheie, MD, PhD, Department of Pathology, Oslo University Hospital, Oslo, Norway, for aiding in this work.
Author Contributions
Author contributions: Guarantors of integrity of entire study, K.E.E., P.D.T., J.K.H.; study concepts/study design or data acquisition or data analysis/interpretation, all authors; manuscript drafting or manuscript revision for important intellectual content, all authors; approval of final version of submitted manuscript, all authors; agrees to ensure any questions related to the work are appropriately resolved, all authors; literature research, K.E.E., M.C.P., J.K.H., L.R.S., T.R.M., O.R.; clinical studies, K.E.E., M.C.P., P.D.T., J.K.H., T.R.M., O.R., A.B.; experimental studies, K.E.E., F.G.Z., L.R.S.; statistical analysis, K.E.E., F.G.Z., L.R.S.; and manuscript editing, K.E.E., M.C.P., F.G.Z., P.D.T., J.K.H., L.R.S., T.R.M., A.B.
References
- 1. . Perfusion MR imaging: moving forward. Radiology 2008;249(2):416–417. Link, Google Scholar
- 2. . Classification of brain tumor type and grade using MRI texture and shape in a machine learning scheme. Magn Reson Med 2009;62(6):1609–1618. Crossref, Medline, Google Scholar
- 3. . Survival analysis of patients with high-grade gliomas based on data mining of imaging variables. AJNR Am J Neuroradiol 2012;33(6):1065–1071. Crossref, Medline, Google Scholar
- 4. . Investigating brain tumor differentiation with diffusion and perfusion metrics at 3T MRI using pattern recognition techniques. Magn Reson Imaging 2013;31(9):1567–1577. Crossref, Medline, Google Scholar
- 5. . Machine learning in preoperative glioma MRI: survival associations by perfusion-based support vector machine outperforms traditional MRI. J Magn Reson Imaging 2014;40(1):47–54. Crossref, Medline, Google Scholar
- 6. . Correlations between perfusion MR imaging cerebral blood volume, microvessel quantification, and clinical outcome using stereotactic analysis in recurrent high-grade glioma. AJNR Am J Neuroradiol 2012;33(1):69–76. Crossref, Medline, Google Scholar
- 7. . Correlation of MR relative cerebral blood volume measurements with cellular density and proliferation in high-grade gliomas: an image-guided biopsy study. AJNR Am J Neuroradiol 2011;32(3):501–506. Crossref, Medline, Google Scholar
- 8. . Cerebral blood volume measurements by perfusion-weighted MR imaging in gliomas: ready for prime time in predicting short-term outcome and recurrent disease? AJNR Am J Neuroradiol 2009;30(4):681–688. Crossref, Medline, Google Scholar
- 9. . Glioma grading by using histogram analysis of blood volume heterogeneity from MR-derived cerebral blood volume maps. Radiology 2008;247(3):808–817. Link, Google Scholar
- 10. . Semi-automated and automated glioma grading using dynamic susceptibility-weighted contrast-enhanced perfusion MRI relative cerebral blood volume measurements. Br J Radiol 2012;85(1020):e1204–e1211. Crossref, Medline, Google Scholar
- 11. . Diffusion-weighted and perfusion MR imaging for brain tumor characterization and assessment of treatment response. Radiology 2006;239(3):632–649. Link, Google Scholar
- 12. . Brain tumors: a multimodality approach with diffusion-weighted imaging, diffusion tensor imaging, magnetic resonance spectroscopy, dynamic susceptibility contrast and dynamic contrast-enhanced magnetic resonance imaging. Magn Reson Imaging Clin N Am 2013;21(2):199–239. Crossref, Medline, Google Scholar
- 13. . 3D Slicer as an image computing platform for the Quantitative Imaging Network. Magn Reson Imaging 2012;30(9):1323–1341. Crossref, Medline, Google Scholar
- 14. . Relative cerebral blood volume maps corrected for contrast agent extravasation significantly correlate with glioma tumor grade, whereas uncorrected maps do not. AJNR Am J Neuroradiol 2006;27(4):859–867. Medline, Google Scholar
- 15. . Automatic vessel removal in gliomas from dynamic susceptibility contrast imaging. Magn Reson Med 2009;61(5):1210–1217. Crossref, Medline, Google Scholar
- 16. . Apparent diffusion coefficient histogram analysis stratifies progression-free and overall survival in patients with recurrent GBM treated with bevacizumab: a multi-center study. J Neurooncol 2012;108(3):491–498. Crossref, Medline, Google Scholar
- 17. . MR imaging predictors of molecular profile and survival: multi-institutional study of the TCGA glioblastoma data set. Radiology 2013;267(2):560–569. Link, Google Scholar
- 18. . Glioblastoma multiforme regional genetic and cellular expression patterns: influence on anatomic and physiologic MR imaging. Radiology 2010;254(2):564–576. Link, Google Scholar
- 19. . Overall survival, prognostic factors, and repeated surgery in a consecutive series of 516 patients with glioblastoma multiforme. Acta Neurol Scand 2010;122(3):159–167. Crossref, Medline, Google Scholar
- 20. . Prognostic factors for survival in adult patients with cerebral low-grade glioma. J Clin Oncol 2002;20(8):2076–2084. Crossref, Medline, Google Scholar
- 21. . Training nu-support vector classifiers: theory and algorithms. Neural Comput 2001;13(9):2119–2147. Crossref, Medline, Google Scholar
- 22. . New support vector algorithms. Neural Comput 2000;12(5):1207–1245. Crossref, Medline, Google Scholar
- 23. . Quantifying tumor vascular heterogeneity with dynamic contrast-enhanced magnetic resonance imaging: a review. J Biomed Biotechnol 2011;2011:732848. Crossref, Medline, Google Scholar
- 24. . Multimodal MR imaging model to predict tumor infiltration in patients with gliomas. Neuroradiology 2014;56(2):107–115. Crossref, Medline, Google Scholar
- 25. . Support vector machine multiparametric MRI identification of pseudoprogression from tumor recurrence in patients with resected glioblastoma. J Magn Reson Imaging 2011;33(2):296–305. Crossref, Medline, Google Scholar
Article History
Received April 1, 2014; revision requested May 8; revision received September 18; accepted October 6; final version accepted October 10.Published online: Dec 8 2014
Published in print: Apr 2015
