Brain tumor segmentation and survival prediction using a cascade of random forests

Lefkovits, Szidónia and Szilágyi, László and Lefkovits, László (2019) Brain tumor segmentation and survival prediction using a cascade of random forests. In: Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries (BrainLes 2018), 16 Sep 2018, Granada, Spain.

Text miccai brats 2018 paper.pdf Restricted to Registered users only Download (2MB)

Abstract

Brain tumor segmentation is a difficult task due to the strongly varying intensity and shape of gliomas. In this paper we propose a multi-stage discriminative framework for brain tumor segmentation based on BraTS 2018 dataset. The framework presented in this paper is a more complex segmentation system than our previous work presented at BraTS 2016. Here we propose a multi-stage discriminative segmentation model, where every stage is a binary classifier based on the random forest algorithm. Our multi-stage system attempts to follow the layered structure of tumor tissues provided in the annotation protocol. In each segmentation stage we dealt with four major difficulties: feature selection, determination of training database used, optimization of classifier performances and image post-processing. The framework was tested on the evaluation images from BraTS 2018. One of the most important results is the determination of the tumor ROI with a sensitivity of approximately 0.99 in stage I by considering only 16% of the brain in the subsequent stages. Based on the segmentation obtained we solved the survival prediction task using a random forest regressor. The results obtained are comparable to the best ones presented in previous BraTS Challenges.

Actions (login required)

Edit Item