A systematic comparison of three-dimensional MR (magnetic resonance) thermography and planning calculations in phantoms for the hyperthermia (HT) SIGMA-Eye applicator. We performed 2 x 6 experiments in a homogeneous cylindrical and a heterogeneous elliptical phantom by adjusting 82 different patterns with different phase control inside an MR tomograph (Siemens Magnetom Symphony, 1.5 Tesla). For MR thermography, we employed the proton resonance frequency shift method with a drift correction based on silicon tubes. For the planning calculations, we used the finite-difference time-domain (FDTD) method and, in addition, modeled the antennas and the transforming network. We generated regions according to a segmentation of bones and tissue, and used an interpolation technique with a subgrid of 0.5 cm size at the interfaces. A Gauss-Newton solver has been developed to adapt phases and amplitudes. A qualitative agreement between the planning program and measurements was obtained, including a correct prediction of hot spot locations. The final deviation between planning and measurement is in the range of 2-3 W/kg, i.e., below 10%. Additional HT phase and amplitude adaptation, as well as position correction of the phantom in the SIGMA-Eye, further improve the results. HT phase corrections in the range of 30-40 degrees and HT amplitude corrections of +/- 20-30% are required for the best agreement. The deviation /MR-FDTD/, and the HT phase/amplitude corrections depend on the type of phantom, certain channel groups, pattern steering, and the positioning error. Appropriate agreement between three-dimensional specific absorption rate distributions measured by MR-thermography and planning calculations is achieved, if the correct position and adapted feed point parameters are considered. As long as feed-point parameters are uncertain (i.e., cannot be directly measured during therapy), a prospective planning will remain difficult. However, we can use the information of MR thermography to better predict the patterns in the future even without the knowledge of feed-point parameters.