We will develop a library of probabilistic tools in the functional language Haskell to deal with probabilistic models such as Markov Random Fields (MRFs). A Markov Random field is a graphical model where each node has a hidden cause, an observable variable which is influenced by the hidden cause and its neighbors, as defined by the graph. This would be used to extend 2D texture generation models (such as Efros and Leung. Texture synthesis by non-parametric sampling. International Conference on Computer Vision (1999)) to 3D in order to generate a model of 3D images of neurons. This model would be useful for tracing out and quantifying the shape and structure of neurons. The model would be tested on images of neurons acquired through a variety of different imaging techniques such as confocal imaging of fluorescent neurons, light microscopy, and two-photon microscopy. Developing a method to understand how microscopy images are generated from hidden sources like this would allow for the automatic tracing and segmentation of a wide range of microscopy images and would could be applied to a number of different research areas, e.g. automated tracing of blood vessels or automating the identification of pathologies from mciroscopy images.
Who am I?:
David Willshaw, Graeme Phillipson, Tom Larkworthy
How is it novel? What is exciting about it?:
Functional languages such as Haskell provide an excellent way to implement mathematical models as they are able to use higher order functions to factorise out common algorithm patterns. As such, a library of the functions which implement an MRF would be useful in enabling people in this area to take advantage of a functional approach when constructing their models.
MRFs have not been applied to the problem of tracing neurons in microscopy images so this approach is novel.
A quick, flexible, accurate and automatic image processing software package will enable biomedical researchers to analyse orders of magnitude more data. In particular this could have a huge impact on the advancement of the field of neuroscience and our understanding of the brain.
Neural reconstruction is such a common problem faced by many neuroscientists that the Allen Institute for Brain Sciences and the Howard Hughes Medical Institute launched the DIADEM (Digital Reconstruction of Axonal and Dendritic Morphology) competition in order to encourage the development of automatic tracing algorithms. The competition organisers have provided 5 datasets including gold standard (manual) reconstructions and a specific metric of success, thus
the problem of tracing in microscopy images is hard but well-defined. More information about the competition can be found at http://www.diademchallenge.org/.
What will I do next? What opportunities will it open up?:
If successful, this technology will be applied to automatic registration and neuron tracing as part of an entry for the DIADEM challenge. This form of analysis can also be applied to medical imaging and other medical research. In future we would like to apply this analysis to fruit fly brain data from our own department, electron microscopy data in collaboration with a Neurobiology group at Dalhousie University, and data collected by Peter Kind’s lab at the Centre for Integrative Physiology (CIP) at Edinburgh University.
Examples of volumetric data that may also benefit from this type of analysis include biomedical imaging such as MRI, fMRI, PET, CT, and CAT scans, and biological microscopy such as Confocal microscopy, Electron Microscopy, 2-Photon Microscopy.
What constitutes success? How risky is it?:
Success will be a library of tools for modelling MRFs in the functional language Haskell. As well as this there will also be a tool for tracing structures in 3D image stacks. Additionally success can be measured by how much faster 3D volumes can be analysed by utilising this technology. Documentation and examples of how to utilise the library from within Haskell will also be provided.
The overall risk for this project is low. The goal of constructing a library for dealing with MRFs in Haskell is easily achievable. The goal of applying it to successfully trace out images of neurons is more risky. However there is reason to be optimistic about success as MRFs have already been applied to modelling natural images. It is likely that MRFs will be more useful in tracing microscopy images than in the analysis of natural images as the image generation in microscopy images is mostly affected by local factors, to which MRFs are better suited, whereas natural images are affected by global factors such as illumination.
What resources do I bring to the project?:
Experience of developing algorithms for 3D machine visions systems, such as the CODA CX-1 (see www.codamotion.com) as well as experience with researching the human visual system, and researching robotic systems. Also there is a library of 3D confocal images available for analysis (see http://fruitfly.inf.ed.ac.uk/braintrap ) and in collaboration with the neuron tracing team we have experience with the DIADEM image dataset and the challenges faced when analysing this type of data.
What resources and expertise do I need?:
One person should be able to complete this investigation. There are several suitable staff available for this project from the ANC. Further benchmarking and analysis work may be carried out in collaboration with other 3D volume analysis groups within the School of Informatics.
What shared resources, if any, will the project create?:
A library of functions in Haskell for dealing with Markov Random Fields. Also a tool for tracing microscopy images of neuron
What is the timescale?:
3 months, starting in January 2010.
2 weeks for creating an MRF library in Haskell.
5 weeks for creating an MRF model of the images of neurons.
2 weeks for testing and bug fixing
2 weeks for data analysis, benchmarking and performance optimisation
1 week for documentation
