Kuijper, ArjanMukhopadhyay, AnirbanFuchs, MoritzMoritzFuchs2022-03-072022-03-072020https://publica.fraunhofer.de/handle/publica/283403Despite being the de-facto standard for medical image segmentation, researchers have identified shortcomings of frequentist U-Nets such as overconfidence and poor out of- distribution generalization. Although their Bayesian counterpart has already been proposed, often these methods rely on the well-known Monte-Carlo Drop Out (MCDO) approximation. We move beyond the MCDO approximation and introduce a novel multi-headed Bayesian U-Net. The proposed approach combines the global exploration capabilities of deep ensembles with the out-of-distribution robustness of Variational Inference. An efficient training strategy, along with an expressive yet general design, ensures superior approximation of the true Bayesian posterior within a reasonable compute requirement. Further we thoroughly analyze different properties of our model and give insights on other prior and regularization techniques. We evaluate our approach on the publicly available BRATS2018 dataset.enLead Topic: Individual HealthResearch Line: Computer vision (CV)medical diagnosismedical imagingdeep learningmachine learning006master thesis