Segmentation of brain tumors and patient survival prediction: Methods for the braTS 2018 challenge

Abstract

Brain tumor localization and segmentation is an important step in the treatment of brain tumor patients. It is the base for later clinical steps, e.g., a possible resection of the tumor. Hence, an automatic segmentation algorithm would be preferable, as it does not suffer from inter-rater variability. On top, results could be available immediately after the brain imaging procedure. Using this automatic tumor segmentation, it could also be possible to predict the survival of patients. The BraTS 2018 challenge consists of these two tasks: tumor segmentation in 3D-MRI images of brain tumor patients and survival prediction based on these images. For the tumor segmentation, we utilize a two-step approach: First, the tumor is located using a 3D U-net. Second, another 3D U-net – more complex, but with a smaller output size – detects subtle differences in the tumor volume, i.e., it segments the located tumor into tumor core, enhanced tumor, and peritumoral edema. The survival prediction of the patients is done with a rather simple, yet accurate algorithm which outperformed other tested approaches on the train set when thoroughly cross-validated. This finding is consistent with our performance on the test set - we achieved 3rd place in the survival prediction task of the BraTS Challenge 2018.