The detection of intracranial aneurysms from 3D time-of-flight magnetic resonance angiography (TOF-MRA) images is a problem of increasing clinical importance. Recently, a streak of methods have shown promising performance by adapting neural networks for 3D semantic segmentation with a patch-based approach. However, these methods are less relevant in a clinical context, where diagnostic decisions depend on detecting objects rather than their segmentation. In particular, evaluation metrics are poorly adapted. In this work, we propose a simple but effective aneurysm detection approach. Based on a 3D adaptation of YOLO architecture, we propose an anchor-free method based on fast data annotation, as well as adapted data sampling and generation strategies to detect aneurysms using a spherical representation. Using 5-fold cross-validation, we compare our method with state-of-the-art nnUNnet and nnDetection baselines on 3 different datasets, including two public ones, with a total of 336 patients. Our method speeds up training and inference time. Based on adapted object detection performance metrics, our method demonstrated comparable or even better performance, with an average précision of 80.44% and a sensitivity of 89.67%, associated with 1.77 false positives per case.