Feature selection and classification using multiple kernel learning for brain tumor segmentation

Abstract

We propose a brain tumor segmentation method from multi-sequence images. The method selects the most relevant features and segments edema and tumor using a classification algorithm based on Multiple Kernel Learning (MKL). Using MKL algorithm, we can associate one or more kernels to each feature. Each kernel is associated to a weight reflecting its importance in the classification. A sparsity constraint on the kernel weights allows to force same weights to be equal to zero corresponding to insignificant kernels (non informative features). Our method was evaluated on real patient dataset of the MICCAI 2012 BraTS challenge. The results show that our method is competitive to the winning methods.