The subject of this talk is the construction of synthetic images of 3-d kidney shapes from real CT images of kidneys. The kidney shapes are a starting point for a more complex longer term project with the ultimate goal of generating a large number of synthetic medical images and shapes for segmentation performance characterisation.

Our data set is of a common type in the statistical analysis of populations of medical images: High Dimension Low Sample Size. For such HDLSS settings classical multivariate analysis methods, such as Principal Component Analysis, are nearly useless, because it is impossible to "sphere the data", since the rank of the covariance matrix is very small compared to the dimension of the data. We adapt the methodology of Independent Component Analysis (ICA) to search for directions of non-Gaussianity in the data. ICA based tests reveal strongly significant non-Gaussian behaviour including apparent outliers.

Further analysis leads to a carefully tuned method for simulating from a population of kidney shapes. ICA based tests on such simulated models show similar results to those obtained for the real data in terms of the deviation from the Gaussian model.

Go back to Statistics seminars

Go to Department of Statistics home page