My research area is image understanding and computer vision, especially in application to medicine and biology. The projects on offer are all related to current research and would be an ideal introduction for students wishing to pursue doctoral studies in medical image analysis. Most projects involve mathematics and the students choosing them will have to be prepared to master the necessary background and techniques - if needed I shall provide support and guidance. I also very much welcome discussions with students interested in medical image analysis who have their own project ideas in this domain.
We have a large database of multispectral images of skin lesions (moles) together with their diagnosis (normal / abnormal and also histological diagnosis). This project seeks a student with interest in machine learning techniques to attempt the abnormality detection directly from the multispectral image data (up to 8-dimensional). Some experience in machine learning is required as I don’t have much expertise in this area; I can provide much guidance on the skin lesions and on the multispectral image data. See sample images
Imagine a set of points (x,y,z) forming a “swiss roll” with values f(x,y,z) increasing from the centre of the roll outwards. Parametrising such sets (i.e. finding a functional form of f) in Cartesian space is quite hard, especially if the geometry of the point set is not known a priori (unlike knowing that the points form a swiss roll). If you could unwrap the roll to form a rectangle g(x’,y’) parametrisation would be trivial as g increases monotonically as a function of (say) y’. Manifold learning is a family of methods based on non-linear dimensionality reduction (see e.g. a Wikipedia page ) that we would like to use to learn a lower-dimensional functional form of abstract models of tissue histology. This could then be used for classification of tissue properties and, further on, for detection of abnormal areas such as early cancer. The project will explore various methods of dimensionality reduction, initially on the swiss roll model.
Macular pigment (MP) plays an important role in maintaining health of a retina. The amount of pigment as well as its distribution can be used for early diagnosis of eye diseases such as Age-related Macular Degeneration (AMD). Our group has developed a novel method for extracting maps showing the distribution of macular pigment from multispectral images (see an example of the map). The challenge remains in deriving quantitative parameters characterising the distribution of MP. Zernike polynomials form an orthogonal basis function set that seems particularly compatible with the kind of shapes we see in MP maps. This project will use optimisation methods to fit Zernike polynomials (ZP) to the MP distribution maps. It is anticipated that the coefficients of ZP will provide a quantitative parametrisation of the macular pigment distribution.
Maintained by Ela Claridge
