A multi-modal data model for morphological segmentation in 3d dosimetry

Patient specific dosimetry established during the last decade in modern radio-therapy. Usually, tracer kinetics in main compartments of observed metabolism is assessed from anterior and posterior whole body scans. The effective doses for each organ, derived by the MIRD scheme, provide evidence for following radiotherapeutic treatment and helps to meet vital dose limits for critical organs, e.g. kidneys. The calculation of individual dose in a three-dimensional context leads to more accurate dose estimates, as was proven by intensive research, but is still on the cusp to clinical application. In this work, a statistical approach, based on multi-modal image and feature data, is presented, to overcome manual segmentation, the most time consuming step, in 3D based dose calculation. 3D data volumes from a hybrid SPECT study, comprising SPECT and CT data, covering main compartments of metabolism, build the image features of a Gaussian classifier. From prior segmentations organspecific membership maps are derived, and substituted as additional feature into the segmentation procedure. Centroids, eccentricity and principal axes of organ models are registered to a rough thresholded image of the SPECT component, and define membership coefficients of the voxels. The new approach yields accurate results, even with real patient data. The new method needs minimal user interaction during selection of some sample regions, thus showing high potential for implementation in a clinical workflow.