Information fusion in biomedical image analysis: combination of data vs. combination of interpretations

Information fusion has, in the form of multiple classifier systems, long been a successful tool in pattern recognition applications. It is also becoming increasingly popular in biomedical image analysis, for example in computer-aided diagnosis and in image segmentation. In this paper, we extend the principles of multiple classifier systems by considering information fusion of classifier inputs rather than on their outputs, as is usually done. We introduce the distinction between combination of data (i.e., classifier inputs) vs. combination of interpretations (i.e., classifier outputs). We illustrate the two levels of information fusion using four different biomedical image analysis applications that can be implemented using fusion of either data or interpretations: atlas-based image segmentation, "average image" tissue classification, multi-spectral classification, and deformation-based group morphometry.