Convolutional Neural Networks for Medical Image Semantic Segmentation

Abstract

This thesis focuses on the problem of medical image segmentation using convolutional neural networks (CNN) with a particular focus on solving the issues that arise in boundary regions and high-frequency areas using different approaches. Convolutional neural networks are recognised by their ability to model local dependencies and hypothesise them to the high-level concepts at the apex of the pyramid of abstractions. Yet, due to the inevitable resolution loss through the network, the inferred label(s) includes little or no locational information. Also, they lack the capacity to model the interdependencies between neighbouring output pixels in a single step training phase. Besides, the lack of enough spatial context information in CNN's sampling frame is an identified issue in medical image semantic segmentation tasks. In this thesis, we aim at addressing the aforementioned shortcomings by modelling the local interdependencies at different levels of abstractions in training and at runtime, preserving and improving context information through the layers, and training the network in a generative adversarial process. The statements are supported by the experimental results that outperform the state of the art solutions for the five medical image semantic segmentation datasets.