Medical image analysis

Digital mammography image simulation

A wide variety of digital mammography systems are now used routinely in breast cancer screening. However these vary greatly in terms of physical performance and cost. The more expensive systems generally have better physical characteristics, but the impact of this on clinical performance is unknown. There is therefore a need to assess the effect of physical characteristics on clinical outcomes. Standard methods of measuring the physical performance of digital detectors are well established. However the relationship between such measures and clinical results are not well understood. Our work uses data and models based on physical measurements to simulate the images of typically used test phantom objects in digital mammography. Cancer Research UK and the Engineering and Physical Sciences Research Council (EPSRC) have awarded a grant of up to £2.5M to the Royal Surrey County Hospital, working with University of Surrey, for a five year programme of research into breast cancer imaging called OPTIMAM. It is hoped this will continue the work to accurately simulate the appearance of breast cancers in digital mammograms, and the measure the performance of radiologists and CAD systems in detecting cancers when different system, radiation doses, beam qualities and image processing are employed.

Collaborator: Royal Surrey Hospital, Guildford, Surrey

Skull stripping

In order to analyse MRI brain data, it is often necessary to strip out those voxels representing non-brain tissue (e.g. skull, sub-cutaneous fat, skin, eyes etc), leaving the 3 key cerebral tissue components: white matter, grey matter and cerebrospinal fluid. We have developed an automatic method of achieving this using a combination of automated region growing, binary morphology and tests have shown that the method produces good quality results in image datasets ranging from adults down to young infants.

Collaborator: Atkinson Morley wing. St George's Hospital NHS Trust, Tooting, London

Partial volume modelling

All digital imaging systems suffer from mixed pixels (in 2D), or in 3D, partial volume voxels or which occur at the boundaries where one object within an image meets another. The finite size of the voxels, together with an imaging system's point spread function, causes mixing of the object intensities. In a biomedical context, this causes loss of contrast, but more importantly leads to difficulties in accurately quantifying tissue (e.g. tumour) volumes. We have developed a number of statistical voxel classifiers aimed at addressing this issue. This has been applied to MR brain data, an example is shown below. Our latest work demonstrates the application of Benford's Law, (an empirical Law used, amongst other things to check fraudulent IRS/tax returns!) to model and quantify the statistical mixing within partial volume voxels. We are currently developing this application using registered PET/CT data, a well as on dynamic MR data.

Collaborators:

  • Atkinson Morley wing. St George's Hospital NHS Trust, Tooting, London
  • Royal Marsden NHS Trust Hospital, Sutton, Surrey
  • Great Ormond Street Hospital for Sick Children, London
  • Atkinson Morley wing. St George's Hospital NHS Trust, Tooting, London
  • Royal Marsden NHS Trust Hospital, Sutton, Surrey
  • Great Ormond Street Hospital for Sick Children, London

Longitudinal segmentation of small structures using non-linear registration

In order to quantify and localise brain atrophy, it is important to be able to segment small structures in the brain, e.g. the hippocampus, on longitudinal series of scans (i.e. scans of the same individual at different time-points). To do this manually is, however, very time-consuming and prone to human error. We have been developing techniques that use non-linear registration to propagate a manual segmentation of the small structure through the series of scans. It is hoped to extend this technique to enable segmentation using a template structure manually segmented on an atlas scan.

Collaborator: Dementia Research Centre, UCL

Wilm's tumour segmentation

Wilm's tumurs are agrressive forms of kidney cancer occurring mostly in young children. Our aim is to develop methods whuich can provide a reliable estimate of the active and total tumour volume for improving dose calculations assoicated with chemotherapy - used to shrink the tumour prior to surgery. Successful segmentation of the necrotic component is relatively straightforward as shown below. This was achieved using a no-parameter region growing tool which only requires initial placement of the seed region. Termination of the region growing process is achieved by dynamically examining the statistics of the resulting histogram of the region. Successul segmentation of the outer active region of the tumour is more challenging, for human expert as well as machine, as the boundaries are poorly defined, and not well constrained. We are currently developing work based on partial volume mixing and active contours to address this issue.

Collaborators: Great Ormond Street Hospital for Sick Children, London

Quantifying and localising calcium deposition in multi-slice CT cardiac images

Quantifying calcium deposition in coronary arteries may be critical to determining if patients are at risk of cardiac artery disease (CAD), and hence need to be prescribed lipid-lowering drugs. Currently, calcium deposits are measured on multi-slice CT (MSCT) images using simple thresholding operations. This gives a single number representing quantification and takes no account of distribution, pattern or intensity of the calcifications. 

This work is focusing on developing methods to analyse the CT images to obtain more information for the clinician about the calcifications present, in particular, providing information on localisation and pattern. The additional analysis methods might include feature-based methods, texture analysis, and multi-resolution analysis. It is envisaged that there will be a large data-set of MSCT images of people who are believed to be at risk of CAD with which to perform this analysis, and which should additionally provide the opportunity for group statistical analysis, for example using PCA methods.

Collaborators:

  • School of Biomedical Sciences, UniS
  • Royal Surrey County Hospital
  • Conquest Hospital, Hastings.

Contact us

Find us

Address
Centre for Vision Speech and Signal Processing
Alan Turing Building (BB)
University of Surrey
Guildford
Surrey
GU2 7XH