“…Image registration accuracy has been investigated for CT to CT liver registration for contrast-enhanced diagnostic CTs [38]. Over the past decade, numerous semi-automatic and automatic approaches for liver segmentation [39,40] on CT that rely on histogrambased methods [41,42], graph cut [43][44][45], region growing [45][46][47], geometric deformable model and level set [48][49][50], probabilistic atlas [51,52], statistical shape models [53][54][55], and recently neural network [56][57][58][59] have been proposed. Despite these efforts, image registration and segmentation remains a challenging task for SIRT application for several reasons: (1) liver is a soft tissue and liver shape is heavily dependent on patient positioning (e.g., the position of the arms); (2) the liver shape in SIRT patients differs from the normal shape, because of preceding treatments (liver resection, liver ablation, chemotherapy) and tumor growth which makes it challenging to use liver segmentation techniques which are dependent on the liver shape for these patients; (3) liver is a soft tissue and its Hounsfield units are similar to those of adjacent organs like the heart, spleen, stomach, and kidney, which makes liver segmentation on non-contrast-enhanced CTs (e.g., CT from MAA study) hard, even for experts; (4) CT from MAA study is not a dedicated diagnostic CT, this low-dose CT usually suffers from streak artifacts; and (5) the interval between the MAA study and the diagnostic high-dose, contrast-enhanced CT from from fluorine-18 fluorodeoxyglucose ( 18 F-FDG) PET/CT study can be up to weeks to even 1 or 2 months and the liver can deform dramatically over time for several reasons, e.g., tumor change.…”