Search

This paper presents a framework that optimizes kernel-based manifold embedding for the characterization of multispectral image data. The hypothesis is that data manifolds corresponding to high-dimensional images can have varying characteristics and types of nonlinearity. As a result, kernel functions must be selected from a wide range of transformations and tuned on an image- and patient-basis. To this end, we introduce a new measure to assess the quality of the kernel transformations that takes into account both local and global relationships in nonlinear manifolds. Furthermore, the calculated measures for each kernel are used to combine the different kernel transformations further highlight the tissue constituents in all regions of the image. Validation with phantom and real multispectral image data shows improvement in the visualization and characterization of the tissue constituents.