A Teacher–Student-based adaptive structural deep learning model and its estimating uncertainty of image data

Abstract

Deep learning has been successfully used as a model that can effectively represent multiple features of the input space and significantly improve image recognition performance on deep architectures. Adaptive structural learning methods of restricted Boltzmann machines (Adaptive RBM) and deep belief networks (Adaptive DBN) have been developed as self-organizing deep learning models. The model uses a neuron generation–annihilation algorithm to find the optimal number of hidden neurons in the RBM for a given input data and then layers a new RBM as a hidden layer on top of the trained RBM to obtain the appropriate DBN structure. The proposed learning model was applied to PET and MRI image data sets in ADNI digital archive for the early detection of MCI (Mild Cognitive Impairment) and AD (Alzheimer's Disease). Two deep learning models were constructed to classify the PET and MRI images, respectively. For the training set, our model showed 99.7% and 99.2% classification accuracy for PET and MRI images, and for the test set, the model showed 98.8% and 96.7% accuracy for them. The Adaptive DBN model achieved the highest classification accuracy among the other CNN models. Moreover, the Teacher–Student-based Adaptive DBN was developed as an ensemble learning to improve the classification power for the test data set of MRI images and the accuracy increased to 98.3%. The accuracy of AD vs CN, MCI vs CN, and MCI vs AD for the MRI images by Teacher–Student-based Adaptive DBN are 98.4%, 98.8%, and 97.8%, respectively. Moreover, the difference between the diagnostic results by the proposed deep learning model and Medical Questionnaire was discussed.