Semantic Segmentation of Brain Tumors in MRI Data Without any Labels

Abstract

Brain MR images are one of the most important instruments for diagnosing neurological disorders such as tumors, infections or trauma. In particular, grade I-IV brain tumors are a well-studied subject for supervised deep learning approaches. However, for a clinical use of these approaches, a very large annotated database that covers all of the occurring variance is necessary. As MR scanners are not quantitative, it is unclear how good supervised approaches, trained on a specific database, will actually perform on a new set of images that may stem from a yet other scanner. We propose a new method for brain tumor segmentation, that can not only identify abnormal regions, but can also delineate brain tumors into three characteristic radiological areas: The edema, the enhancing core, and the non-enhancing and necrotic tissue. Our concept is based on FLAIR and T1CE MRI sequences, where abnormalities are detected with a variational autoencoder trained on healthy examples. The detected areas are finally postprocessed via Gaussian Mixture Models and finally classified according to the three defined labels. We show results on the BraTS2018 dataset and compare these to previously published unsupervised segmentation results as well as to the results of the BraTS challenge 2018. Our developed unsupervised anomaly detection approach is on par with previously published methods. Meanwhile, the semantic segmentation - a new and unique model - shows encouraging results.