One of the most challenges in medical imaging is the lack of data and annotated data. It is proven that classical segmentation methods such as U-NET are useful but still limited due to the lack of annotated data. Using a weakly supervised learning is a promising way to address this problem, however, it is challenging to train one model to detect and locate efficiently different type of lesions due to the huge variation in images. In this paper, we present a novel approach to locate different type of lesions in positron emission tomography (PET) images using only a class label at the image-level. First, a simple convolutional neural network classifier is trained to predict the type of cancer on two 2D MIP images. Then, a pseudo-localization of the tumor is generated using class activation maps, back-propagated and corrected in a multitask learning approach with prior knowledge, resulting in a tumor detection mask. Finally, we use the mask generated from the two 2D images to detect the tumor in the 3D image. The advantage of our proposed method consists of detecting the whole tumor volume in 3D images, using only two 2D images of PET image, and showing a very promising results. It can be used as a tool to locate very efficiently tumors in a PET scan, which is a time-consuming task for physicians. In addition, we show that our proposed method can be used to conduct a radiomics study with state of the art results.