Rotate to view desktop version ...

Image Segmentation

Deep Learning for Segmentation
Image Segmentation is traditionally performed using (probabilistic) atlases. This process can be very time consuming as prior normalization of intensities as well as non-linear spacial registration to the atlas is required. Deep-Learning, however, can be used for the rapid segmention of anatomical structures - often within seconds - and promises to provide scalable approaches for processing of large datasets (such as the ADNI, UK-Biobank, Rhineland Study etc) and can provide fast decision support in clinical workflows. In this domain, our contributions focus on the development of novel convolutional neural-network architectures, as well as on the rigorous validation of these methods with respect to reliability, sensitivity and generalizability.
FatSegNet - Dixon MRI Body Fat
In [1] we introduce a competitive dense network for the segmentation of various fat compartments from Dixon MRI in the Rhineland Study. The challenge here was to work with only a very limited training set (30 subjects with mutliple 2D slices) and still generalize to population data.
Brain MRI processing pipelines, such as FreeSurfer, are highly complex software packages that perform various processing steps. Often the processing speeds of 6-8 h per image prevents them from being applicable to very large cohort studies. We have therefore developed a neural network (FastSurferNet) that can segment a full 3D brain MRI via view-aggreation into close to 100 structures within 1 minute. Furthermore, building on FreeSurfer we speed-up surface processing to deliver a FreeSurfer alternative with only 1.5 h processing time. The corresponding paper [2] is currently under review and can be found at arXiv FastSurfer.
Selected Publications (full list):
Martin Reuter - MIT - Cambridge, MA, USA - EMail:
Last modified 14.01.2022 -- [xhtml] -- [css] -- [welcome]