Multi-plane UNet plus plus Ensemble for Glioblastoma Segmentation

Glioblastoma multiforme (grade four glioma, GBM) is the most aggressive malignant tumor in the brain and usually treated by combined surgery, chemo- and radiotherapy. The O-6-methylguanine-DNA methyltransferase (MGMT) promoter methylation status was shown to be predictive of GBM sensitivity to alkylating agent chemotherapy and is a promising marker for personalized treatment. In this paper we propose to use a multi-plane ensemble of UNet++ models for the segmentation of gliomas in MRI scans, using a combination of Dice loss and boundary loss for training. For the prediction of MGMT promoter methylation, we use an ensemble of 3D EfficientNet (one per MRI modality). Both, the UNet++ ensemble and EfficientNet are trained and validated on data provided in the context of the Brain Tumor Segmentation Challenge (BraTS) 2021, containing 2.000 fully annotated glioma samples with four different MRI modalities. We achieve Dice scores of 0.792, 0.835, and 0.906 as well as Hausdorff distances of 16.61, 10.11, and 4.54 for enhancing tumor, tumor core and whole tumor, respectively. For MGMT promoter methylation status prediction, an AUROC of 0.577 is obtained.