Unsupervised deep neural scheme for mobile phone based unlabelled medical image classification

Abstract

The ever-increasing size of medical datasets combined with increasing numbers of missing data has made unsupervised learning one of the strong focus of significant practical importance in the real world. There has been much interest in applying unsupervised techniques that incorporate information from unlabeled data for higher representation. Accurate identification of diseases within a short span is of great importance due to the global increase in new disease cases. Medical diagnosis by using automated computer-aided procedure is more effective compared to the manual pathological methods. This study presents an automatic identification of infected erythrocytes parasites and intestinal parasites using a new deep learning method. This new deep neural network architecture consists of autoencoder followed by support vector machine. The entire network consists of two phases: in the first phase the autoencoder takes the network weights with their initial values by unsupervised greedy layer-wise technique, and the support vector machine in the second phase are fine-tuned by the backpropagation algorithm. Our extensive experimental results demonstrate that the proposed deep neural network can obtain better performance in terms of accuracy and time than other broadly used deep learning techniques.