Brain Tumor Segmentation Using Neural Ordinary Differential Equations with UNet-Context Encoding Network

Glioblastoma Multiforme (GBM) are the most aggressive brain tumor types and because of their heterogeneity in shape and appearance, their segmentation becomes a challenging task. Automated brain tumor segmentation using Magnetic Resonance Imaging (MRI) plays a key role in disease diagnosis, surgical planning, and brain tumor tracking. Medical image segmentation using deep learning-based U-Net architectures are the state-of-the-art. Despite their improved performance, these architectures require optimization for each segmentation task. Introducing a continuous depth learning with context encoding in deep CNN models for semantic segmentation enable 3D image analysis quantifications in many applications. In this work, we propose Neural Ordinary Differential Equations (NODE) with 3D UNet-Context Encoding (UNCE), a continuous depth deep learning network for improved brain tumor segmentation. We showed that these NODEs can be implemented within the U-Net framework to improve segmentation performance. This year we participated for the Brain Tumor Segmentation (BraTS) continuous evaluation and our model was trained using the same MRI image sets of RSNA-ASNR-MICCAI Brain Tumor Segmentation (BraTS) Challenge 2021. Our model is evaluated on unseen holdout data included i) the BraTS 2021 Challenge test data, ii) SSA adult patient populations of brain diffuse glioma (Africa-BraTS), and iii) from another independent pediatric population of diffuse intrinsic pontine glioma (DIPG) patients. The mean DSC for the BraTS test dataset are: 0.797797 (ET), 0.825647 (TC) and 0.894891 (WT) respectively. For the Africa-BraTS dataset the performance of our model improves which indicating the generalizability of our model to new, out-of-sample populations for adult brain tumor cases.