Detecting cerebral aneurysms is an important clinical task of brain computed tomography angiography (CTA). However, human interpretation could be time consuming due to the small size of some aneurysms. In this work, we proposed DeepBrain, a deep learning based cerebral aneurysm detection and localization algorithm. The algorithm consisted of a 3D faster region-proposal convolution neural network for aneurysm detection and localization, and a 3D multi-scale fully convolutional neural network for false positive reduction. Furthermore, a novel hierarchical non-maximum suppression algorithm was proposed to process the detection results in 3D, which greatly reduced the time complexity by eliminating unnecessary comparisons. DeepBrain was trained and tested on 550 brain CTA scans and achieved sensitivity of 93.3% with 0.3 false positives per patient on average.