Sparse Classification with MRI Based Markers for Neuromuscular Disease Categorization
In this paper, we present a novel method for disease classification between two patient populations based on features extracted from Magnetic Resonance Imaging (MRI) data. Anatomically meaningful features are extracted from structural data (T1- and T2-weighted MR images) and Diffusion Tensor Imaging (DTI) data, and used to train a new machine learning algorithm, the k-support SVM (ksup-SVM). The k-support regularized SVM has an inherent feature selection property, and thus it eliminates the requirement for a separate feature selection step. Our dataset consists of patients that suffer from facioscapulohumeral muscular dystrophy (FSH) and Myotonic muscular dystrophy type 1 (DM1) and our proposed method achieves a high performance. More specifically, it achieves a mean Area Under the Curve (AUC) of 0.7141 and mean accuracy 77%±0.013. Moreover, we provide a sparsity visualization of the features in order to indentify their discriminative value. The results suggest the potential of the combined use of MR markers to diagnose myopathies, and the general utility of the ksup-SVM. Source code is also available at https://gitorious.org/ksup-svm.
Keywordsmyopathies Diffusion Tensor Imaging k-support regularized SVM
Unable to display preview. Download preview PDF.
- 11.Argyriou, A., Foygel, R., Srebro, N.: Sparse Prediction with the k-Support Norm. In: Advances in Neural Information Processing Systems, NIPS (2012)Google Scholar
- 14.Wang, P., Gur, R., Verma, R.: A novel framework for identifying dti-based brain patterns of schizophrenia. In: ISMRM, Toronto, pp. 3–9 (2008)Google Scholar