Luengo, Imanol
(2018)
Hierarchical super-regions and their applications to biological volume segmentation.
PhD thesis, University of Nottingham.
Abstract
Advances in Biological Imaging technology have made possible imaging of sub-cellular samples with an unprecedented resolution. By using Tomographic Reconstruction biological researchers can now obtain volumetric reconstructions for whole cells in near-native state using cryo-Soft X-ray Tomography or even smaller sub-cellular regions with cryo-Electron Tomography. These technologies allow for visualisation, exploration and analysis of very exciting biological samples, however, it doesn’t come without its challenges. Poor signal-to-noise ratio, low contrast and other sample preparation and re-construction artefacts make these 3D datasets to be a great challenge for the image processing and computer vision community. Without previous available annotations due to the biological relevance of the datasets (which makes them not being publicly available) and the scarce previous research in the field, (semi-)automatic segmentation of these datasets tends to fail.
In order to bring state-of-the-art in computer vision closer to the biological community and overcome the difficulties previously mentioned, we are going to build towards a semi-automatic segmentation framework. To do so, we will first introduce superpixels, a group of adjacent pixels that share similar characteristics that reduce whole images to a few superpixels that still preserve important information of the image. Superpixels have been used in the recent literature to speed up object detection, tracking and scene parsing systems. The reduced representation of the image with a few regions allows for faster processing on the subsequent algorithms applied over them. Two novel superpixel algorithms will be presented, introducing with them what we call a Super-Region Hierarchy. A Super-Region Hierarchy is composed of similar regions agglomerated hierarchically. We will show that exploiting this hierarchy in both directions (bottom-up and top-down) helps improving the quality of the superpixels and generalizing them toimages of large dimensionality.
Then, superpixels are going to be extended to 3D (named supervoxels), resulting in a variation of two new algorithms ready to be applied to large biological volumes. We will show that representing biological volumes with supervoxels helps not only to dramatically reduce the computational complexity of the analysis (as billions of voxels can be accurately represented with few thousand supervoxels), but also improve the accuracy of the analysis itself by reducing the local noisy neighbourhood of these datasets when grouping voxel features within supervoxels. These regions are only as powerful as the features that represent them, and thus, an in-depth discussion about biological features and grouping methods will lead the way to our first interactive segmentation model, by gathering contextual information from super-regions and hierarchical segmentation layers to allow for segmentation of large regions of the volume with few user input (in the form of annotations or scribbles).
Moving forward to improve the interactive segmentation model, a novel algorithm will be presented to extract the most representative (or relevant) sub-volumes from a 3D dataset, since the lack of training data is one of the deciding factors for automatic approaches to fail. We will show that by serving small sub-volumes to the user to be segmented and applying Active Learning to select the next best sub-volume, the number of user interactions needed to completely segment a 3D volume is dramatically reduced. A novel classifier based on Random Forests will be presented to better benefit from these regions of known shape.
To finish, SuRVoS will be introduced. A novel fully functional and publicly available workbench based on the work presented here. It is a software tool that comprises most of the ideas, problem formulations and algorithms into a single user interface. It allows a user to interactively segment arbitrary volumetric datasets in a very intuitive and easy to use manner. We have then covered all the topics from data representation to segmentation of biological volumes, and provide with a software tool that hopefully will help closing the gap between biological imaging and computer vision, allowing to generate annotations (or ground truth as it is known in computer vision) much quicker with the aim of gathering a large biological segmentation database to be used in future large-scale completely automatic projects.
Actions (Archive Staff Only)
 |
Edit View |
Tools
Tools
![[thumbnail of Thesis.pdf]](https://eprints.nottingham.ac.uk/style/images/fileicons/application_pdf.png "Thesis.pdf")