Abstract
Background and Objective Cerebral microbleeding (CMB) is associated with many brain diseases, such as dementia and vascular disease. CMBs can be detected by brain magnetic resonance imaging (MRI). Susceptibility weighted imaging (SWI) is commonly employed since it can give better sensitivity than standard MRI. However, CMBs are usually small and they can be distributed throughout brain, manual analysis is arduous and tedious. We proposed to use deep learning methods to detect CMBs. First, we collected 64 brain SWI. We used a sliding window size of 61x61 pixel to generate 10000 samples. Then, we labeled the samples as non-CMB or CMB manually. Finally, we employed convolutional neural network (CNN) for classification. Results In the experiment, we used 8000 samples to train the CNN, the rest 2000 for testing. The proposed method yielded a sensitivity of 97.29%, a specificity of 92.23%, and an overall accuracy of 96.05%. Conclusions The results suggested our method can detect and locate CMBs automatically and accurately.