Project Suggestions 2012/13
___________________________


Here are some projects suggestions for this year.  Have a look at my
home page http://www.cs.ucl.ac.uk/staff/d.alexander or my group's
website http://cmic.cs.ucl.ac.uk/mig to get an idea of other work
going on; I am happy to discuss projects in any of those areas too.


Models of dispersion in capillary networks
==========================================

The project will construct and evaluate mathematical and computational
models of the MRI signal from water molecules in blood flowing through
capillary networks.  It contributes to the broader research program of
the Microstructure Imaging Group: cmic.cs.ucl.ac.uk/mig.
Diffusion-weighted MRI holds great promise as a non-invasive
diagnostic probe in cancer imaging.  Blood flow in capillaries (the
intra-voxel incoherent motion or IVIM effect) is a significant
confound that complicates its practical application.  However,
modelling the effect potentially allows us to exploit
diffusion-weighted imaging in cancer more effectively and may even
provide additional information that is currently discarded.  The
project takes the first steps towards this by constructing candidate
mathematical models and validating them against computer simulations.

With Becky Shipley (UCL Dept. Mechanical Engineering)


Cortical grey matter segmentation using diffusion MRI
=====================================================

Segmenting the cortex of the brain in MRI images in a consistent and
reliable way is a key challenge in neuroscience and neurology, because
it enables meaningful comparisons of brain function across
individuals.  For example, to make statements about how two
individuals execute a particular cognitive task, we need to be able to
say with confidence which parts of the brain are active during the
task.  Recently we have derived new measurements from diffusion MRI
that appear to separate the brain into functionally distinct areas.
The aim of this project is to start to exploit this new information by
combining those measurements with state of the art unsupervised
learning and clustering algorithms to provide a brain segmentation
with more precision and complexity than achievable by any existing
approach.

With Zoltan Nagy and Marty Sereno from UCL's Institute of Neurology


Image-based diagnosis of multiple sclerosis
===========================================

This project will use classification techniques from machine learning
to make diagnoses of multiple sclerosis from brain images.  UCL's
institue of neurology has large databases of images from
multiple-sclerosis patients and normal controls that we can use to
train classifiers.  More accurate diagnosis from imaging will have a
direct impact on quality of life of these patients by expediting
decisions about which treatments they should receive.  Previous work
has managed to get around 80% correct diagnosis using machine learning
techniques, but the features the algorithms learn from are manually
defined so expensive to obtain.  This project will look at learning
from much more directly obtainable image features aiming for a truly
automated computer assisted diagnosis system.

With Olga Ciccarelli from UCL's Institute of Neurology


Stochastic optimization inspired by biological foraging models
==============================================================

Stochastic optimization techniques, such as simulated annealing,
genetic algorithms, self-organizing migratory algorithms (SOMA), and
differential evolution, avoid problems of local minima that confound
gradient descent algorithms.  They are often the only viable
techniques for minimizing high dimensional functions with complex
topology.  Many of these algorithms maintain a population of candidate
solutions, which migrate over the search space in various ways.  For
example, the SOMA algorithm uses an analogy of herding cattle, which
migrate towards rich areas of pasture.  The precise mechanism of the
migration determines the efficiency of the algorithm and the SOMA
strategy proves consistently effective in a diverse range of problems.
The aim of this project is to test a range of alternative biologically
inspired migration strategies to improve the efficiency of the search.


Cancer grading from histology
=============================

This project aims to automate the process of cancer grading using
image processing, computer vision and machine learning techniques.
The standard way to diagnose cancer, and to determine how malignant a
particular tumour is, is through biopsy and histology: a small piece
of tumour is extracted from the patient; a pathologist looks at it
under a microscope and decides, from the cellular make-up of the
tissue, the nature of the tumour and thus the appropriate treatment.
We will aim here to automate the pathologist's process of grading the
cancer from the microscope image.  We have lots of data to train and
test classifiers.  A key part of the project however will be to
determine what are the most important image features required to
maximize classification performance.

With Laura Panagiotaki from UCL's Centre for Medical Image Computing