Semi-Automatic Segmentation for Serial Section Electron Microscopy Images.
Independent thesis Advanced level (degree of Master (Two Years)), 20 credits / 30 HE creditsStudent thesis
Providing the detailed 3D structure of the full set of neurons and synapses in an organism is a fundamental problem in neuroscience. For performing such 3D reconstruction at nanoscale, serial section Electron Microscopy has been proven as the only existing technique. However, this technique produces Terabytes of imaging data which analyzing those images manually is an extremely time consuming task. To reconstruct each single neuron, one needs to manually annotate and track the neuron through a stack consisting of thousands of images. This manual annotation constitutes the main bottleneck in the 3D reconstruction workflow. In this thesis we have introduced a novel approach to reduce the human effort needed in this technique. We use Graph-Cut to compute segmentation. We compute segmentations based on the image intensities and the output of predictors for low-level features, such as cell membrane and internal cell structures. We have extended the standard Graph-Cut to incorporate information from neighboring layers to make the segmentation consistent between layers.
We evaluated our method on serial section images of Pristionchus pacificus, a nematode similar to C.elegans. The dataset contains approximately 4,000 layers, each with a size of 20K by 20K pixels. We tested our method to segment neurons on a subset of the whole stack without any human intervention. We found that the accuracy of the generated segmentations on each layer was comparable to human performance. Furthermore, we successfully tracked and segmented one neuron through the whole image stack with only a few manual corrections. As run-time was a critical issue, we investigated in detail the trade-off between speed and accuracy on various levels of our processing pipeline. Our proposed segmentation method is implemented as a plug-in to Fiji, a commonly used software package for biological image analysis.
Place, publisher, year, edition, pages
Trita-CSC-E, ISSN 1653-5715 ; 2012:045
IdentifiersURN: urn:nbn:se:kth:diva-130987OAI: oai:DiVA.org:kth-130987DiVA: diva2:654433
Master of Science - Computational and Systems Biology