Multi-Atlas Based Pseudo-CT Synthesis Using Multimodal Image Registration and Local Atlas Fusion Strategies

The synthesis of pseudo-CT images is of particular interest in the development of hybrid PET-MRI devices and MRI-guided radiotherapy. These images can be used for attenuation correction during PET image reconstruction. Furthermore, using MRI-based radiotherapy planning would enable a more accurate dosimetry planning due to the superior soft tissue contrast of the scans. The previously proposed methods for pseudo-CT synthesis are characterised by mainly two drawbacks. First, most proposed methods are limited to the head and neck region and therefore not feasible in case of whole body applications. Second, the presence of aligned training pairs of both MRI and CT scans for a number of subjects is assumed. In this work, we present preliminary results for atlas-based approaches using multiple CT atlas scans (from different patients) to synthesise a pseudo-CT image for a new patient using only their MRI data. This application requires accurate and robust deformable multimodal registration. We employed a recent discrete optimisation registration framework together with a self-similarity-based metric to accurately match the CT atlases to the anatomy of the patient. The registered atlases are then jointly combined by means of local fusion strategies. We apply our method to different 3D whole body MRI scans and a total of 18 3D whole body CT atlases. In addition to intensity fusion, the proposed methods can also be used for label fusion. Since evaluation based directly on synthesised intensity values is problematic, we use the Dice overlap after the fusion of segmentation labels as a proxy measure. Our proposed new method, which uses MIND descriptors for multimodal label fusion shows overall the best results.