Hierarchical Bayesian classification of multimodal medical images

Abstract

It has gradually been recognised that Bayesian algorithms are more widely applicable and reliable than ad hoc algorithms. Advantages include the use of explicit and realistic stochastic models making it easier to understand the working behind the algorithm and allowing confidence statements about conclusions. The authors propose a method, within a Bayesian framework, to assimilate information from images obtained from different modalities at different resolutions. The algorithm is used with a pair of images, from which a fused high resolution image and improved data reconstructions are simultaneously obtained. The authors illustrate their method by 2 examples, the first fuses a pair of SPECT and CT phantom images and the second a pair of MR brain scan images, obtained from different acquisition techniques. The authors provide a pseudo-comparison of the latter example with a commercially available package called ANALYZE. However, the phantom images from physical experiment given here provide a true validation and performance of the model.