Positron Emission Tomography (PET) is a Nuclear Medicine technique with a wide range of applications, particularly on Oncology. The imminent clinical introduction of Long Axial Field-of-View (LAFOV) scanners has raised the interest of the community, due to its enhanced sensitivity and lower doses required. In this project, we aim to develop a novel framework on Monte Carlo techniques, in which a set of Total-Body simulated images is used to train networks for organ and lesion segmentation. Our methodology can be divided into phantom generation, simulation and segmentation. We established a distribution of parameters to obtain map cohorts with anatomical variability, which were simulated using the open-source platform SimPET and finally segmented at organ and lesion levels using U-Net architectures. Usually, PET images are segmented together with Computerized Tomography (CT), integrated on PET/CT scans for attenuation correction. Direct PET segmentation might be useful in the cases of low-quality CT, in novel attenuation correction approaches without using anatomical imaging and specifically for non-anatomical lesions. Our model gives excellent overlapping results on most of the organs among simulated test images. The model also presents good performance on lesion segmentation, if they are large and well-contrasted enough. Furthermore, our method overcomes the sparsity of data limitation and automatically annotates the image database by using the activity maps information.