Tumor volume and heterogeneity are important for patient diagnosis, and automatic lesion segmentation is needed to compute this information from routine CT-Scans. Training a supervised neural network to solve these tasks demands good quality annotations on a large quantity of fully annotated scans, which are difficult and time-consuming to obtain. We propose a fast automatic method using morphological operators to create 3D masks from hand drawn contours of the lesions on their largest axial slice. This type of annotation leads to more precise 3D masks than points or ellipses. Thus, the obtained mask may be used to train end-to-end neural networks for detection and semantic segmentation of lesions on CT-Scans in 3D. We tested this methodology on the LIDC-LUNA dataset to produce the 3D masks from automatically selected 2D annotations. We also produced 3D masks of 115 lung lesions from their 2D contours, and compared them to ground truth 3D masks on an in-house dataset. The results are promising, and the method could be adapted to other organs.