Advancing efficiency and robustness of neural networks for imaging

Abstract

Enabling machines to see and analyze the world is a longstanding research objective. Advances in computer vision have the potential of influencing many aspects of our lives as they can enable machines to tackle a variety of tasks. Great progress in computer vision has been made, catalyzed by recent progress in machine learning and especially the breakthroughs achieved by deep artificial neural networks.

Goal of this work is to alleviate limitations of deep neural networks that hinder their large-scale adoption for real-world applications. To this end, it investigates methodologies for constructing and training deep neural networks with low computational requirements. Moreover, it explores strategies for achieving robust performance on unseen data. Of particular interest is the application of segmenting volumetric medical scans because of the technical challenges it imposes, as well as its clinical importance. The developed methodologies are generic and of relevance to a broader computer vision and machine learning audience.

More specifically, this work introduces an efficient 3D convolutional neural network architecture, which achieves high performance for segmentation of volumetric medical images, an application previously hindered by high computational requirements of 3D networks. It then investigates sensitivity of network performance on hyper-parameter configuration, which we interpret as overfitting the model configuration to the data available during development. It is shown that ensembling a set of models with diverse configurations mitigates this and improves generalization. The thesis then explores how to utilize unlabelled data for learning representations that generalize better. It investigates domain adaptation and introduces an architecture for adversarial networks tailored for adaptation of segmentation networks. Finally, a novel semi-supervised learning method is proposed that introduces a graph in the latent space of a neural network to capture relations between labelled and unlabelled samples. It then regularizes the embedding to form a compact cluster per class, which improves generalization.