Abstract
1338
Objectives The automatic analysis and detection of subtle changes between patient images over time or between multispectral data sets is an extremely important tool both in patient diagnosis and in follow up studies when determining treatment response. The purpose of this study is to investigate the application of local kernel regression as a robust and automatic technique for determining such changes.
Methods The method is based on the computation of a local kernel from a spatially registered reference image and that of a target image containing potential significant differences or changes. This kernel is used as a feature and compared against homologous features from the target image. Local regression kernels robustly derive local geometric structures of images by analyzing the pixel(voxel) differences based on estimated gradients and use this structural information to determine the shape and size of a canonical kernel. The local steering kernel is modelled by a multivariate Gaussian with parameters N, the number of pixels(voxels) in a local window around the current sampling point xi. The associated covariance matrix Ci modifies the shape and size of the local kernel reflecting the image’s intrinsic local geometry. To provide a similarity metric the cosine of the angle between the two vectors of samples derived from the traget and reference images was derived and functions as the Pearson correlation coefficient for mean centred data.
Results We present examples of the application of LKR to a planar data set of 12 renal failure patients suffering from tertiary hyperparathyroidism. Imaging was performed in dual isotope mode using I-123 as the reference thyroid image and Tc99m MIBI as the target image i.e. thyroid + adenoma.
Conclusions Local kernel regression is an effective and robust approach to detecting subtle structural differences between image data sets in nuclear medicine