Brain tumor segmentation from multiple MRI sequences using multiple kernel learning

We propose a brain tumor segmentation method from multi-spectral MRI images. First, a large set of features based on wavelet coefficients, is computed on all types of images for a small number of voxels, allowing us to build a training feature base which is not homogeneous due to different types of image. The segmentation task is then viewed as a learning problem where only the most significant features from the feature base should be selected and then a classifier can be used. The new idea is to use Multiple Kernel Learning (MKL) by associating one or more kernels to each feature in order to solve jointly the two problems: selection of the features and their corresponding kernels and training of the classifier. All types of images are then segmented using the trained classifier on the selected features. Our algorithm was tested on the real data provided by the challenge of Brats 2012 and was compared to the resulting top methods. The results show good performance of our method.