This paper presents a 3D nonlinear PET-CT image registration method guided by a B-Spline Free-Form Deformations (FFD) model, dedicated to thoracic and abdominal regions. It is divided into two stages: one FFD-based registration of structures that can be identified in both images; and a whole-image intensity registration step constrained by the FFD computed during the first step. Different similarity criteria have been adopted for both stages: Root Mean Square (RMS) to register recognized structures and Normalized Mutual Information (NMI) for optimizing the whole-image intensity stage. Structure segmentation is performed according to a hierarchical procedure, where the extraction of a given structure is driven by information derived from a simpler one. This information is composed of spatial constraints and expressed by the means of regions of interest, in which a 3D simplex mesh deformable model based method is applied. The results have been very positively evaluated by three medical experts.