The recent progresses performed in imaging, computational and technological fields bring new opportunities to achieve high precision radiation dose delivery. However, IMRT requires a particular attention at the target delineation step to avoid inadequate dosage to TVs/OARs. In this context, the biological information provided by PET might advantageously complete CT-Scan to refine the target delineation in HNSCC and lung cancer. Integrating PET into the treatment planning however requires the use and validation of accurate and reproducible segmentation methods, which adequately integrate the PET image properties such as the blur effect and the high level of noise. In this context, we developed specific tools, i.e. edge-preserving filters for denoising and deconvolution algorithms for deblurring that allowed the detection of gradient intensity peaks. Our gradient-based method has been validated on phantom and patient materials, and proved to be more accurate than threshold-based approaches. With this tool in hand, we demonstrated that the use of FDG-PET resulted in smaller TVs than the CT-based TVs, on both pre- and per-treatment images, and significantly improved the dose distributions to the TVs/OARs. This opens avenues for dose escalation strategies that might potentially improve the tumor local control.