Unpaired cross-modal translation with cyclic loss is being increasingly used for a large variety of medical imaging applications such as e.g. segmentation. However, finer-scale details like tumors may be lost during translation, which is a critical limitation in oncological imaging. In this paper, we propose to address the problem of vanishing tumors for cross-modal segmentation. First, we propose a new method to insert realistic tumors in 3-D images using a deep generative model trained on a single 2-D image. Second, we leverage the proposed model using a new unpaired-then-paired two-stage I2I architecture to better penalize the suppression of tumors in cross-modal segmentation. In our experiments, we validate our model on the ongoing MICCAI crossMoDa tumor segmentation challenge, where we demonstrate superior performance over CycleGAN-based models.