For magnetic resonance imaging no protocol dependent intensity standard, like the Hounsfield units in computed tomography, is available due to magnetic field inhomogeneities in both B0 and RF excitation fields. One type of variation is that intensities of the same tissue class differ throughout a single volume. In order to deal with that problem, a variety of algorithms for bias field correction were developed in the last decade. However, these methods do not solve the other type of problem: a certain measured intensity cannot be associated with a tissue class.  For segmentation, a missing protocol dependent standard intensity scale has the disadvantage that for every new suspect an individual training of the used (statistical) model has to be performed. For this reason the clinical applicability of many algorithms is low due to runtime restrictions. Furthermore, visualization systems cannot use standard presets (e.g., transfer functions) to visualize certain organs or tissue classes. The settings have to be adjusted for every single scan. Hence, a second class of approaches dealing with inter-scan intensity standardization was developed by several authors. State-of-the-art algorithms, generally, standardize the observed intensities using a single image at a time and ignore spatially adjacent images. For many applications this is sufficient, because in many regions of the body a gray value in one image is associated with exactly one intensity in another sequence (e.g., the brain). In general, however, this is not the case.

The algorithms developed in this project utilize all acquired images for intensity standardization. With that, it is possible to separately correct tissue classes that have the same intensity in one image but can be distinguished using more data sets. Furthermore, the introduced approach does not rely on any assumptions about the shape of the joint histograms used. Thus the method is completely independent from the application, region of interest (brain, thorax, pelvis, etc.), scanning protocol (e.g., T1-, T2-weighted) as long as there are learned histograms available for the task.