Representation learning of brain diseases using deep learning with various datasets

Abstract

For the last decade, deep neural networks known as “deep learning”, have demonstrated “super-human” performance in some complex areas such as computer vision including medical imaging, and autonomous driving, due to their large model capacity and effective feature extraction from “big data”. Depending on detailed design of neural network models, it is possible to extract salient features effectively, or “representation learning”, to predict the crucial factors with high reliability in the field of clinical medicine. In this thesis, to contribute to the conquest of brain diseases such as brain tumors and depression, which had limited improvement in treatment outcomes, we study representation learning of various brain diseases using various datasets with development and validation of an appropriate design of neural network models, to deal with effective feature extraction, prediction of important indicators, and analysis of saliency of extracted features.