Deep learning approaches for classification and semantic segmentation of medical image data

Abstract

In this dissertation, we address the deep-learning approaches for classification and semantic segmentation of medical image data. We propose a Transferable Ranking Convolutional Neural Network to consider the inter-class relationship, and applied it to glaucoma detection in fundus images. Also, we propose an encoder-decoder in encoder-decoder Convolutional Neural Network architecture where precise segmentation is possible from the beginning of decoding by transmitting all levels of features extracted from every encoder block, and applied it to three types of main vessel segmentation in coronary angiography. The proposed classification and semantic segmentation methods showed higher classification accuracy and Dice Similarity Coefficient score compared to existing methods and it can be effectively applied to other medical image data with similar data characteristics.