Deep random forest-based learning transfer to SVM for brain tumor segmentation

Using neuroimaging techniques to diagnose brain tumors and detect both visible and invisible cancer cells infiltration boundaries motivated the emergence of diverse tumor segmentation algorithms. Noting the large variability in both tumor appearance and shape, the task of automatic segmentation becomes more difficult. In this paper, we propose a random-forest (RF) based learning transfer to SVM classifier method for segmenting tumor lesions while capturing their complex characteristics. Our framework is composed of two cascaded stages. In the first stage, we train a random forest to learn the mapping from the image space to the tumor label space. In the testing stage, we use the predicted label output from the random forest and feed it along with the testing intensity image to an SVM classifier to get the refined segmentation. Then we make our RF-SVM cascaded classification steps deep through an iterative process. We tested our method on 20 patients with high-grade gliomas from the Brain Tumor Image Segmentation Challenge (BRATS) dataset. Our proposed framework significantly outperformed SVM-based segmentation and RF-based segmentation-when used solely.