Deep Neural Network Applications for Medical Imaging Diagnostics

Abstract

Medical imaging is a good tool for diagnosing disease, but it is impossible to diagnose disease without medical interpretation. For medical interpretation, skilled clinicians are required, but to train clinicians take over a decade and million dollars. In this dissertation, we propose automated diagnostic system based on a deep neural network as a solution for this problem. The first part of this thesis deals with an automatic diagnostic method based on the deep extreme learning machine. In this chapter, we propose a modified model of the deep extreme learning machine for multi-modal features. This model extracts optimal sparse hierarchical multi-modality features for automatic diagnosis of Alzheimer's disease from magnetic resonance imaging, positron emission tomography and cerebrospinal fluid. The second part of this study is about automatic staging of rectal cancer. In this topic, we constructed the deep neural network model that simulates the doctor’s criteria for determing stage of rectal cancer. The model consists of three independent deep neural networks, two segmentation models and a classification model. Our proposed model shows a significantly higher classification rate for rectal cancer staging than the previously proposed models. We expect that the proposed algorithms will be useful for automated diagnostic systems in the near future.