Abstract: Cerebral Microbleeds (CMBs) are small chronic brain hemorrhages which are caused by structural abnormalities of the small vessels of the brain. CMBs are increasingly found in various patient populations and disease settings, including first-ever and recurrent ischaemic or haemorrhagic stroke, Alzheimer's disease, vascular cognitive impairment and healthy elderly individuals. Previous clinical routine, CMBs are manually detected by radiologists and it may produce the errors and time-consuming. In this paper, we proposed a two-stage cascaded framework to detect CMBs from magnetic resonance (MR) images using 3D convolutional neural network (CNN). We first endeavour a 3D fully convolutional Network (FCN) technique to recover the candidates with high probabilities of being CMBs, and afterward apply a well-trained 3D CNN discrimination model to recognize CMBs from hard mimics. Within our framework, the Hierarchical Centroid Shape Descriptor (HCSD) is allows to select only those having a specific structure. We illustrate that proposed framework can be utilized to prepare efficient and accurate classifiers that could introduce further Computer-aided diagnosis.

Keywords: Cerebral microbleeds, convolutional neural networks, Hierarchical Centroid Shape Descriptor, MRI.