CASUS Institute Seminar, Martin Jones, Deputy Head of Microscopy Prototyping in the Electron Microscopy STP at the Francis Crick Institute, London, UK
Volume electron microscopy techniques in bioimaging, such as serial block face SEM (SBF SEM), focused ion beam SEM (FIB SEM) and array tomography (AT), produce datasets frequently reaching the terabyte regime. In many cases, much of the analysis is still performed manually, which represents a major bottleneck in the workflow, restricting the amount of information that can be extracted from these rich datasets. A very common task is “segmentation”, where objects of interest are delineated from the background. Modern computational techniques such as deep learning have shown a great deal of promise for automating many complex segmentation tasks. However, a major limitation is the lack of availability of the prerequisite “ground truth” training data which is often painstakingly generated manually by researchers. Borrowing a strategy developed by astronomers, Martin’s “Etch a Cell” project on the Zooniverse online citizen science platform invites volunteers from all walks of life to contribute annotations. By aggregating this non-expert data, his team is able to produce robust training data that has been used to train a convolutional neural network with no requirement for expert segmentations at all.
A further problem is the generalization of these trained models to new datasets with minimal additional manual interaction. Martin will show that pre-trained models are able to efficiently adapt to new segmentation tasks on different datasets.
Martin’s talk was originally planned for early April but had to be postponed at short notice. We apologize for the inconvenience caused and hope to welcome you at the new date.
The event is organized with a videoconferencing tool by Zoom Inc. If you want to join the talk remotely, please ask for the login details via contact@casus.science.