Cerebral Microbleeds Detection via Convolutional Neural Network with and Without Batch Normalization

Abstract

Cerebral microbleeds (CMBs) as a kind of subclinical sign are associated with cerebrovascular and cognitive diseases, as well as normal aging. Hence, it is important to detect CMBs automatically and accurately for helping medical research and preventing related diseases. CMBs can be visualized as small and rounded radiological entities via susceptibility-weighted imaging (SWI). So far, some advances about detecting CMBs automatically have been achieved. In this study, a designed CNN structure is for further improving the performance of detecting CMBs automatically. Furthermore, a breakthrough technique named batch normalization (BN), which is widely used in training deep neural networks for accelerating the training process, was tested. The performance of CNN with BN was compared to that without BN. It is found that the latter model achieved better prediction results. Afterward, four state-of-the-art methods were compared to the designed CNN. The comparison shows the designed CNN achieved the best performance with a sensitivity of 99.69%, a specificity of 96.5%, and an accuracy of 98.09%.