Automatic segmentation of brain tumors using tensor analysis and multimodal MRI

Glioma detection and classification is an critical step to diagnose and select the correct treatment for the brain tumours. There has been advances in glioma research and Magnetic Resonance Imaging (MRI) is the most accurate non-invasive medical tool to localize and analyse brain cancer.The scientific global community has been organizing challenges of open data analysis to push forward automatic algorithms to tackle this task. In this paper we analyse part of such challenge data, the Multimodal Brain Tumor Image Segmentation Benchmark (BRATS), with novel algorithms using partial learning to test an active learning methodology and tensor-based image modelling methods to deal with the fusion of the multimodal MRI data into one space. A Random Forest classifier is used for pixel classification. Our results show an error rates of 0.011 up to 0.057 for intra-subject classification. These results are promising compared to other studies. We plan to extend this method to use more than 3 MRI modalities and present a full active learning approach.