This work deals with the use of a probabilistic quad-tree graph (Hidden Markov Tree, HMT) to provide fast computation, improved robustness and an effective interpretational framework for image analysis and processing in oncology. Thanks to two efficient aspects (multi observation and multi resolution) of HMT and Bayesian inference, we exploited joint statistical dependencies between hidden states to handle the entire data stack. This new flexible framework was applied first to mono modal PET image denoising taking into consideration simultaneously the Wavelets and Contourlets transforms through multi observation capability of the model. Secondly, the developed approach was tested for multi modality image segmentation in order to take advantage of the high resolution of the morphological computed tomography (CT) image and the high contrast of the functional positron emission tomography (PET) image. On the one hand, denoising performed through the wavelet-contourlet combined multi observation HMT led to the best trade-off between denoising and quantitative bias compared to wavelet or contourlet only denoising. On the other hand, PET/CT segmentation led to a reliable tumor segmentation taking advantage of both PET and CT complementary information regarding tissues of interest. Future work will investigate the potential of the HMT for PET/MR and multi tracer PET image analysis. Moreover, we will investigate the added value of Pairwise Markov Tree (PMT) models and evidence theory within this context.