Cerebral Microbleeds Detection via Discrete Wavelet Transform and Back Propagation Neural Network

Cerebral microbleeds (CMBs) are small perivascular hemosiderin deposits leaked through cerebral small vessels in normal or near normal tissue. The positions distribution of CMBs can indicate some underlying aetiologies. CMBs can be visualized by susceptibility-weighted imaging (SWI) which is high sensitivity to hemosiderin. In this paper, we proposed a hybrid method to detect the CMBs automatically. This method first applied discrete wavelet transform (DWT) to extract the features of the brain images, and then employed principal component analysis (PCA) to perform reduction of features. At last, the obtained features were inputted to back propagation shallow neural network (BPNN) with a single-hidden layer for training and prediction. K-fold cross validation was applied to avoid overfitting and evaluate the generalization ability of BPNN for selecting the best model. Based on this method, a good result was obtained with a sensitivity of 88.47 +/- 0.96%, a specificity of 88.38 +/- 1.00%, and an accuracy of 88.43 +/- 0.97%, which is better than two state-of-the-art approaches.