3D Magnetic Resonance Angiography (MRA) plays a vital role in the detection of Unruptured Intracranial Aneurysms (UIAs). In recent years, deep learning-based detection of UIAs from 3D MRA scans has gained popularity due to its high efficiency and accuracy in diagnosis. However, training a robust 3D deep learning model usually requires large amounts of real-world data, which is not easy to obtain due to privacy concerns. An intuitive solution to limited real-world training data is to generate synthetic data from limited real-world samples. The generated 3D volumes, however, need to have ground truth masks for use during model training. Unfortunately, labeling aneurysm masks manually is a time-consuming and non-trivial task. In this paper, we investigate the feasibility of generating 3D MRA volumes from limited real-world 3D MRA scans and ground truth masks using a segmentation-guided (or ground truth-guided) approach based on modern generative models. We evaluate the quality of the generated volumes at the latent space level and investigate the impact of this approach on aneurysm detection models.