Random decision forests for automatic brain tumor segmentation on multi-modal MRI images

Brain tumour segmentation from Magnetic Resonance Imaging (MRI) scans have an important role in the early tumour diagnosis and radiotherapy planning. However, MRI images of the brain contain complex characteristics, such as high diversity in tumour appearance and ambiguous tumour boundaries, even when using multi-sequence MRI images. We propose a fully automatic segmentation algorithm based on a Random Decision Forest, using a k-fold cross-validation approach. The extracted features are the intensity complemented with other appearance and context based features. The post-processing phase has a morphological filter to deal with misclassification errors. Our method is capable of detecting the tumour and segmenting the different tumorous tissues of the glioma achieving competitive results.