The aim of this thesis is to develop an approach for the automatic generation of "patient-specific" 3D anatomical models of the respiratory system, based on a generic reference patient "an Atlas", avoiding the manual segmentation of certain organs "such as the diaphragm". The aim is to compute, extract and apply geometric transformations to map the generic "Atlas" model onto a new patient, exploring deep learning methods and GNNs (Graph Neural Networks). We also aim to develop methods for deforming 3D surfaces (from the atlas to a new patient), while preserving the quality of models adapted to physical simulations, and the constraints of contacts between organs. The models generated will be used to model and simulate the respiratory system using a patient-specific biomechanical approach, in order to quantify lung deformations and predict the position and trajectory of lung tumors with a high degree of accuracy.