Automated Brain Image Classification Based on VGG-16 and Transfer Learning

The last few decades have witnessed active research in the domain of pathological brain image classification starting from classical to the deep learning approaches like convolutional neural networks(CNN). The classical machine learning methods need hand-crafted features to perform classification. CNN's, on the other hand, performs classification by extracting image features directly from raw images. The features extracted by CNN strongly depends on the training data set size. If the size is small, CNN tends to overfit. So, deep CNN's(DCNN) with transfer learning has evolved. The prime aim of the present paper is to explore the capability of a pre-trained DCNN VGG-16 model with transfer learning for pathological brain image categorization. Only, the last few layers of the VGG-16 model were replaced to accommodate new image categories in the present application. The pre-trained model with transfer learning has been validated on the dataset taken from the Harvard Medical School repository, comprising of normal as well as abnormal MR images with different neurological diseases. The data set was then partitioned using a 10-fold cross-validation mechanism. The validation on the test set using sensitivity (Se), specificity (Sp), and, accuracy (Acc), reveal that the pre-trained VGG-16 model with transfer learning exhibited the best performance in contrast to the other existing state-of-the-art works. Moreover, the approach provides categorization with an end-to-end structure on raw images without any hand-crafted attribute extraction.