Deep Learning for Computer-aided Diagnosis of Brain Diseases Through MRI Multi-classification

Methods for automated and accurate classification of brain magnetic resonance (MR) images have been widely proposed for clinical medicine and biomedical research. In this paper, we presented a hybrid computer-aided diagnosis (CAD) system based on the following techniques: median filter for de-noising, region growing algorithm for segmentation, discrete wavelet transformation (DWT) for feature extraction and dropout convolutional neural network (Dropout-CNN) for classification. Our study differentiated normal brain and four types of diseases, which was different from any published studies that only classified the brain MR images into normal and abnormal categories. Furthermore, we introduced the term of singularity to evaluate the confusion where a disease misdiagnosed as other types. As the result, our proposal produced 100% sensitivity rate, 100% specificity rate, 91.7% accuracy rate and 8.7% singularity rate. In comparison with former methods, our results showed that Dropout-CNN gave the highest sensitivity and specificity as well as the second lowest singularity rate. Additionally, it had better performance in infectious disease, normal brain, and neoplastic disease than some published methods.