doi:10.1016/j.neuroimage.2007.04.009 
Copyright © 2007 Elsevier Inc. All rights reserved.
Multivariate examination of brain abnormality using both structural and functional MRI
Yong Fana, 
, 
, Hengyi Raob, Hallam Hurtc, Joan Giannettac, Marc Korczykowskib, David Sherac, Brian B. Avantsa, James C. Geea, Jiongjiong Wangb and Dinggang Shena, ,
aDepartment of Radiology, University of Pennsylvania, PA 19104, USA
bDepartment of Neurology, Center for Functional Neuroimaging, University of Pennsylvania, PA 19104, USA
cDepartment of Pediatrics, Division of Neonatology, The Children’s Hospital of Philadelphia, PA 19104, USA
Received 25 October 2006;
revised 2 April 2007;
accepted 10 April 2007.
Available online 19 April 2007.
References and further reading may be available for this article. To view references and further reading you must
purchase this article.
Abstract
A multivariate classification approach has been presented to examine the brain abnormalities, i.e., due to prenatal cocaine exposure, using both structural and functional brain images. First, a regional statistical feature extraction scheme was adopted to capture discriminative features from voxel-wise morphometric and functional representations of brain images, in order to reduce the dimensionality of the features used for classification, as well as to achieve the robustness to registration error and inter-subject variations. Then, this feature extraction method was used in conjunction with a hybrid feature selection method and a nonlinear support vector machine for the classification of brain abnormalities. This brain classification approach has been applied to detecting the brain abnormality associated with prenatal cocaine exposure in adolescents. A promising classification performance was achieved on a data set of 49 subjects (24 normal and 25 prenatally cocaine-exposed teenagers), with a leave-one-out cross-validation. Experimental results demonstrated the efficacy of our method, as well as the importance of incorporating both structural and functional images for brain classification. Moreover, spatial patterns of group difference derived from the constructed classifier were mostly consistent with the results of the conventional statistical analysis method. Therefore, the proposed approach provided not only a multivariate classification method for detecting brain abnormalities, but also an alternative way for group analysis of multimodality images.
Fig. 1. A framework for brain image classification. In particular, there are three components in the feature extraction step, ① brain template space partition, ② regional feature extraction, and ③ statistical and compact representation of regional features. The finally constructed classifier in the training stage is used to classify a new testing sample as described in the testing panel.
Fig. 2. Illustration of the estimation of the spatial consistency of a feature f located at u (white dot) in the feature map. The spatial consistency of a feature is estimated from training samples by the intraclass correlation coefficient among all features in its spatial neighborhood. The feature values within the given feature's spatial neighborhood constitute the rows of a data matrix, whereas the feature values of same spatial location of all training samples constitute the columns of the data matrix. Based on the data matrix, the intraclass correlation coefficient can be computed by the method in McGraw and Wong (1996). The mathematical formulation is detailed in the Appendix.
Fig. 3. The change of classification rate with respect to the different number of both structural and functional features used for classification and the different size of kernel used in SVM.
Fig. 4. ROC curve of the classifier that yields the best classification rate. Numbers on the curve are the correct classification rates (%). The area under the ROC curve is 0.91.
Fig. 5. The change of classification rate with respect to the different number of structural features used for classification and the different size of kernel used in SVM.
Fig. 6. The change of classification rate with respect to the different number of functional features used for classification and the different size of kernel used in SVM.
Fig. 7. Group differences identified by our pattern classification method for functional feature map (left is left). The significance of group difference in each region is color-coded according to the color bar shown in the bottom, with 1 as relatively the most important for classification.
Fig. 8. Group differences identified by our pattern classification method for GM feature map (left is left). The significance of group difference in each region is color-coded according to the color bar shown in the bottom, with 1 as relatively the most important for classification.
Fig. 9. Group differences identified by our pattern classification method for WM feature map (left is left). The significance of group difference in each region is color-coded according to the color bar shown in the bottom, with 1 as relatively the most important for classification.
Table 1.
Comparison on different feature extraction methods in brain classification, using the proposed hybrid feature selection and nonlinear SVM
The best classification rates obtained by the methods are reported, respectively.
Table 2.
Comparison on different feature extraction methods in brain classification, using the ranking-based feature selection and linear SVM
The best classification rates obtained by the methods are reported, respectively.
Abstract
Purchase PDF (1466 K)
E-mail Article
Add to my Quick Links
Cited By in Scopus (2)
Purchase PDF (1105 K)
