EXPLORE PUBLICATIONS BY COUNTRIES

Cerebral MRI, Tumor, Segmentation, Feature Selection, Multiclass, Classification, Multiple Kernel Learning, Multimodal.

We propose a brain tumor segmentation method from multi-spectral MRI images. The method is based on classification and uses Multiple Kernel Learning (MKL) which jointly selects one or more kernels associated to each feature and trains SVM (Support Vector Machine).

First, a large set of features based on wavelet decomposition is computed on a small number of voxels for all types of images. The most significant features from the feature base are then selected and a classifier is then learned. The images are segmented using the trained classifier on the selected features. In our framework, a second step called re-learning is added. It consists in training again a classifier from a reduced part of the training set located around the segmented tumor in the first step. A fusion of both segmentation procures the final results.

Our algorithm was tested on the real data provided by the challenge of Brats 2012. This dataset includes 20 high-grade glioma patients and 10 low-grade glioma patients. For each patient, T1, T2, FLAIR, and post-Gadolinium T1 MR images are available. The results show good performances of our method.