Rückblick auf die BVM 2021 an der OTH Regensburg

Vorsitz: Heinz Handels, Christoph Palm

PREISTRÄGER: Fabian Isensee, Deutsches Krebsforschungszentrum (DKFZ), Abteilung Medizinische Bildverarbeitung

From Manual to Automated Design of Biomedical Semantic Segmentation Methods

The dataset dependency of current deep learning based semantic segmentation methods in the biomedical domain has severe consequences for the field. Not only does there exist no out-of-the-box tool that can be used by non-experts to get access to state-of-the-art segmentation for their custom dataset, but the progress in methodological research is severely hampered as well. Because methods  are  only  compatible  with  a  narrow  selection  of  similar  datasets  and  the dataset size is rather small (several hundreds of raining cases at best) researchers struggle to discern noise from true methodological improvements.  This thesis aims at breaking the dataset dependency of current segmentation methods. First, we develop and analyze four state-of-the-art segmentation methods targeting  three  different  segmentation  tasks:  brain  tumor  segmentation,  cardiac substructure segmentation and kidney and kidney tumor segmentation. Second, we use the lessons learned from these methods to construct nnU-Net, the first dataset-agnostic segmentation method for 3D segmentation of biomedical images.  nnU-Net analyzes each dataset it is being applied to and automatically adapts its entire pipeline, including the network architecture, to match it.  Despite  its  generic  nature,  nnU-Net  sets  a  new  state  of  the  art  in  29  out  of  49 segmentation tasks across 19 diverse datasets while closely matching state-of-the-art performance in the remainder. We make nnU-Net available to the community.  As an out-of-the-box segmentation tool it makes state-of-the-art segmentation algorithms accessible to anybody.  As a framework it catalyzes future method development by enabling researchers to run a principled evaluation of their method across multiple diverse datasets.

Dauer: 15 min


We use cookies to improve the usability of the website. By visiting you agree to this.