Computer Analysis of 3D MR images of the Brain

Introduction
With the advent of non-invasive visualisation techniques, such as Magnetic Resonance Imaging, has come a need for new tools to process and interpret the results. Given a 3D image of the living brain and the right tools, diagnosticians could do their work as early as possible and neurosurgeons could plan operations more fully in advance. In addition a much larger number of subjects could be available for study, advancing knowledge boundaries, and enabling comparisons within and between groups.

The right tools should be able to locate specific structures, label areas within them, and quantify or classify the results of such a search.

Point Distrubution Models (PDM)
Point Distribution Models are statistical models which can
record the basic shape of an object plus its observed variation over a training set. They are built by placing points on the same significant landmarks on each example, and statistically analysing their 3D co-ordinates over the set. With this knowledge, plus stored attribute information about the neighbourhood around each point, the model can search through a new image to find an example of the object.

Modelling the Brain
The sulcal fissures of the cortical surface provide significant anatomical landmarks and these were chosen for the basis of a PDM. Simple image processing can be used to automatically locate the appropriate points over the sulci (see figure above), however, this gives differing numbers of points between individuals and no correspondence between them. In
order to build a PDM the points must be matched by using available information, such as the configuration of neighbours, or the local surface curvature. This point matching is one of the main issues currently being addressed.

Results
Initial attempts to correspond the point sets have already generated models which exhibit sensible properties, and have yielded some preliminary insights into the cortical shape and pattern changes over a set of normal individuals. These models can be used to examine a new 3D image for the locations of the sulci, and improving the accuracy of the search is another ongoing issue.

Applications
The uses of such a model can be broadly split into three categories:

Descriptive:
The model itself provides insights into the variations over a training set. For example a model built from a time series of images for one patient will provide information about the progress of a disease. Alternatively, building a model over a group of subjects provides details of the characteristics of that group. This could provide insights into the physical symptoms accompanying diseases such as Alzheimer's or Schizophrenia. Also, the characteristics of different groups can be compared. This might be different diseases, or sex, or handedness.

Search:
Having located the fissures, these can be automatically labelled and/or measured, to identify or monitor specific changes in a patient. Similarly, from their location the cortical surface can be extracted and also labelled, measured, or visualised. Again this might be useful for diagnosis or for surgical planning.

Classification:
If several models for different groups are available then the model of best fit gives a classification (diagnosis) for a particular example. This might enable early identification of problems or indications of treatments.

Angela Caunce:

Research Courses Clinical Radiology Industrial liaison Contact Us Join Us Search Home

Top Page