Thesis: Stray neutron doses are currently not evaluated during treatment planning within proton therapy centers since treatment planning systems (TPS) do not allow this feature while Monte Carlo (MC) simulations and measurements are unsuitable for routine practice. The PhD aims at developing an analytical model dedicated to the estimation of stray neutron doses to healthy organs which remains easy-to-use in clinical routine. First, the existing MCNPX model of the gantry installed at the Curie institute – proton therapy center of Orsay (CPO) was extended to three additional treatment configurations (energy at the beam line entrance of 162, 192 and 220 MeV). Then, the comparison of simulations and measurements was carried out to verify the ability of the MC model to reproduce primary proton dose distributions (in depth and lateral) as well as the stray neutron field. Errors within 2 mm were observed for primary protons. For stray neutrons, simulations overestimated measurements by up to a factor of ~2 and ~6 for spectrometry and dose equivalent in a solid phantom, respectively. The result analysis enabled to identify the source of these errors and to put into perspective new studies in order to improve both experimental measurements and MC simulations. Secondly, MC simulations were used to calculate neutron doses to healthy organs of a one-year-old, a ten-year-old and an adult voxelized phantoms, treated for a carniopharyngioma. Treatment parameters were individually varied to study their respective influence on neutron doses. Parameters in decreasing order of influence are: beam incidence, organ-to-collimator and organ-to-treatment field distances, patient’ size/age, treatment energy, modulation width, collimator aperture, etc. Based on these calculations, recommendations were given to reduce neutron doses. Thirdly, an analytical model was developed complying with a use in clinical routine, for all tumor localizations and proton therapy facilities. The model was trained to reproduce calculated neutron doses to healthy organs and showed errors within ~30% and ~60 µGy Gy⁻¹ between learning data and predicted values; this was separately done for each beam incidence. Next, the analytical model was validated against neutron dose calculations not considered during the training step. Overall, satisfactory errors were observed within ~30% and ~100 µGy Gy⁻¹. This highlighted the flexibility and reliability of the analytical model. Finally, the training of the analytical model made using neutron dose equivalent measured in a solid phantom at the center Antoine Lacassagne confirmed its universality while also indicating that additional modifications are required to enhance its accuracy.