[en] One of the main factors of error for semi-quantitative analysis in positron emission tomography (PET) imaging for diagnosis and patient follow up, as well as new flourishing applications like image guided radiotherapy, is the methodology used to define the volumes of interest in the functional images. This is explained by poor image quality in emission tomography resulting from noise and partial volume effects induced blurring, as well as the variability of acquisition protocols, scanner models and image reconstruction procedures. The large number of proposed methodologies for the definition of a PET volume of interest does not help either. The majority of such proposed approaches are based on deterministic binary thresholding that are not robust to contrast variation and noise. In addition, these methodologies are usually unable to correctly handle heterogeneous uptake inside tumours. The objective of this thesis is to develop an automatic, robust, accurate and reproducible 3D image segmentation approach for the functional volumes determination of tumours of all sizes and shapes, and whose activity distribution may be strongly heterogeneous. The approach we have developed is based on a statistical image segmentation framework, combined with a fuzzy measure, which allows to take into account both noisy and blurry properties of nuclear medicine images. It uses a stochastic iterative parameters estimation and a locally adaptive model of the voxel and its neighbours for the estimation and segmentation. The developed approaches have been evaluated using a large array of datasets, comprising both simulated and real acquisitions of phantoms and tumours. The results obtained on phantom acquisitions allowed to validate the accuracy of the segmentation with respect to the size of considered structures, down to 13 mm in diameter (about twice the spatial resolution of a typical PET scanner), as well as its robustness with respect to noise, contrast variation, acquisition parameters, scanner models or reconstruction algorithms. The performance of the developed algorithm is shown to be superior to thresholding reference methodologies. The results demonstrate the ability of the developed approach to accurately delineate tumours with complex shapes and activity distributions, for which the reference methodologies fail to generate coherent segmentation maps. The algorithm is also able to delineate multiples regions inside the tumour, whereas reference methodologies are usually binary only. Both robustness and accuracy results demonstrate that the proposed methodology may be used in clinical context for diagnosis and patients follow up, as well as for radiotherapy treatment planning and 'dose painting', facilitating optimized dosimetry and potentially reduced doses delivered to healthy tissues around the tumour and nearby organs. Such studies to evaluate the impact of the methodology in radiotherapy treatment planning have already started in a project which aims to explore the potential of the algorithm which has been successfully patented. (author)