Search Contact information
University of Cambridge Home Department of Engineering
University of Cambridge > Engineering Department > Machine Intelligence Lab > Medical Imaging Group

Graham Treece

[ Introduction | Teaching | Research | Publications | Software | Videos | Personal ]

Click here for other medical imaging projects offered by Graham Treece.

F-GMT11-3: Smoothing of surface-mapped data for hip fracture analysis.

The thickness of the cortex (outer bone surface) can be measured at lots of locations over the femoral head: this is crucial information for determining risk of fracture. These measurement locations are at the vertices of a polygonal mesh which covers the femoral surface. We can display the measurements on this surface using a colourmap. Some of them are quite precise, others are really noisy. If we want to look at statistics over lots of femurs, we have to smooth the data to remove the noise. But by how much? This data is smoothed a little. This data is smoothed a lot. What we really need is to smooth a different amount at each location to reduce the measurement noise to the same level.

Hip fracture is a major issue affecting millions of people annually. Hip fracture risk is mainly related to the strength and thickness of the denser layer of cortical bone surrounding the cancellous bone in the centre. It is known that this layer of bone can become thinner with increasing age, particularly in regions of the femur which are often involved in fracture due to a fall.

We have recently developed a technique which can accurately estimate cortical thickness and map it to a 3D computer surface of the hip. We have been using this across large patient cohorts to try to work out which bits of the cortical layer are most critical. To do this we need to measure values from lots of patients and study them all mapped to a common femoral surface. But some of the measurements are quite noisy, and we have to smooth them to increase the statistical power. Ideally we want this smoothing to vary over the surface according to how good the measurements are. We already have a fast method - Non-Parametric Regressions (NPR) - which can do this sort of non-uniform smoothing on image data (see below).

This project will investigate the possibility of applying non-uniform data smoothing with NPR on data irregularly distributed across a meshed surface (see above). There are all sorts of issues related to this. How can we calculate data derivatives across the surface? How can we speed up the computation? The matrix techniques used for image smoothing don't apply directly to surface smoothing.

This is an algorithmic development / computational geometry / software project, so experience of writing software is essential, though the development could also be done using Matlab. The project involves some collaboration with the bone research group at Addenbrooke's Hospital.

Here are some measurements (actually of tissue stiffness) which have been taken at lots of locations on a regular grid. Just like with cortical thickness measurements, we can also work out how noisy they are: black is good precision, white is very poor. With this sort of regularly spaced data, we can use Non-Parametric Regression to smooth the data making use of the noise estimates: the bad measurements end up getting smoothed more. We can still control the overall extent of smoothing - this sets the precision of the smoothed data which is now the same everywhere as a result of the smoothing process.
© 2005 Cambridge University Engineering Dept and Graham Treece .
Information provided by gmt11