Automatic Segmentation of Brain Tumor Parts from MRI Data Using a Random Forest Classifier

The segmentation of brain tumor and the separation of its parts like the enhancing core or edema represents a highly important problem, since a fine solution offers precise diagnosis and better opportunities in radiotherapy planning or follow-up studies after interventions. Brain tumor segmentation is also a highly challenging task, due to the wide variety of lesion appearances, the possible presence of noise effects, and the differences in MRI scanner sensitivity. This paper is a preliminary study of a random forest (RF) based solution for the tumor part segmentation problem using multi-spectral MRI data. The proposed method is trained and tested using the 220 high-grade glioma records of the BraTS 2015 train data set. These records are preprocessed to eliminate noise effects and to generate 100 additional features to the four observed ones. The output of the RF classifier is fed directly to statistical evaluation, in order to investigate the direct contribution of the RF to the accurate segmentation. The overall Dice scores exceeding 82% for the whole tumor, 80% for the enhancing core, 74% for the tumor core, and 72% for the edema, make the random forest classifier a good candidate to be successful as the core of a multistage brain tumor part segmentation procedure.