Bermudez-Chacon Roger, Marquez-Neila Pablo, Salzmann Mathieu, Fua Pascal (2018), A domain-adaptive two-stream U-Net for electron microscopy image segmentation, in 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018)
, Washington, DCIEEE, Washington DC, USA.
MosinskaAgata, Marquez NeilaPablo, KosinskiMateusz, FuaPascal (2018), Beyond the Pixel-Wise Loss for Topology-Aware Delineation, in Medical Image Computing and Computer Assisted Intervention
, Granada, SpainSpringer International Publishing, Cham.
Glowacki Przemysaw, Pinheiro Miguel, Mosinska Agata, Türetken Engin, Lebrecht Daniel, Sznitman Raphael, Holtmaat Anthony, Kybic Jan, Fua Pascal (2018), Reconstructing Evolving Tree Structures in Time Lapse Sequences by Enforcing Time-Consistency, in Transactions on Pattern Analysis and Machine Intelligence
, 40(3), 755-761.
MosinskaAgata, TarnawskiJakub, FuaPascal (2017), Active Learning and Proofreading for Delineation of Curvilinear Structures, in Medical Image Computing and Computer-Assisted Intervention
, Springer, Quebec City, Canada.
Mosinska Agata, Sznitman Raphael, Głowacki Przemysław, Fua Pascal (2016), Active Learning for Delineation of Curvilinear Structures Agata, in Computer Vision and Pattern Recognition
, IEEE, Las Vegas, NE, USA.
Bemudez Roger, Becker Carlos, Salzmann Matthieu, Fua Pascal (2016), Scalable Unsupervised Domain Adaptation for Electron Microscopy, in Medical Image Computing and Computer Assisted Intervention
, MICCAI, Athens, Greece.
Imaging modalities such as Electron and Light Microscopy can now deliver high-quality, high-resolution image stacks of neural structures. Manual and semi-automated segmentation or annotation tools are then used to extract structures of interest. However, while the ever growing mass of available imagery should help unlock the secrets of neural function, the required amounts of human intervention remain a major bottleneck. Therefore, there has been great interest in automating the annotation process and most state-of-the-art algorithms rely on Machine Learning.However, such algorithms still require significant amounts of manual annotation for training purposes. For everyday scenes, this can be done by crowd-sourcing. In microscopy, this is not an option because only experts whose time is scarce and precious can annotate them reliably. This is problematic because the data preparation processes tend to be complicated and not easily repeatable, which means that a classifier trained on one acquisition will not perform very well on a new one, even when using the same modality.In earlier work, we have investigated approaches that involve acquiring sufficient amounts of labeled training data after one specific image acquisition and then using it in conjunction with a small amount of additional labeled training data that can be acquired quickly after each subsequent one to retrain the classifiers. We have so far focused mostly on synapse detection and delineation of curvilinear structures. In the proposed project continuation, we intend to broaden the scope of this effort and to develop generally applicable Domain Adaptation that are generally applicable in the field of bio-medical image processing.