Detecting cerebral microbleeds with transfer learning

Cerebral microbleeds (CMBs) are small perivascular hemosiderin deposits leaked from cerebral small vessels in normal (or near normal) tissue. It is important to detect CMBs accurately and reliably for diagnosing and researching some cerebrovascular diseases and cognitive dysfunctions. In the last decade, several approaches based on traditional machine learning and classical convolutional neural network (CNN) were developed for detecting CMBs semi-automatically and automatically. In recent years, numerous advanced variants of CNN with deeper structure have been developed for image recognition, showing better performances comparing with classical CNN. In particular, ResNet proposed recently won the championships on many important image recognition benchmarks because of its extremely deep representations. In view of this, we proposed a method based on ResNet-50 for exploring the possibility of further improving the accuracy of CMBs detection in this study. Due to our small CMB samples size, transfer learning was employed. Based on the transfer learning of ResNet-50, we achieved a high performance with a sensitivity of 95.71 +/- 1.044%, a specificity of 99.21 +/- 0.076%, and an accuracy of 97.46 +/- 0.524% in format of average +/- standard deviation, which outperformed three state-of-the-art methods.