Fast DRR Generation with Light Fields

Computer-aided treatment planning greatly improves the therapeutical index in radiotherapy. However, treatment plans need to be registered with the actual patient position on the treatment device before therapeutical irradiation can be started. Registration is commonly carried out by comparing a set of 2D control radiographs of the patient, the so-called portal images (PI), with 2D synthetic radiographs calculated from the 3D CT scan data with which the treatment plan was established. Computing these digitally reconstructed radiographs (DRR) is an important step in intensity-based PI to CT image registration methods. However, classic volume rendering algorithms are time consuming which turns the DRR generation step into a bottle neck of the registration process. The light field method known from Computer Graphics was recently proposed as a technique to precompute a part of the DRR generation process. We studied the suitability of the different existing light field parameterizations for the registration process and present two novel approaches that are optimized for the radiotherapy context. A geometric analysis of the projection space is carried out that makes it possible to significantly reduce the light field storage space requirements. A novel ray approximation error estimation is introduced which helps to find the best light field configuration in terms of rendering quality for a given projection space. Finally, a novel discretisation of the DPP light field is proposed which makes fast rendering possible while view dependent biases are significantly reduced.