Public defence, Computer Science, MSc Joel Honkamaa

Title of the thesis: Advances in Medical Image Registration with Neural Networks

Thesis defender: Joel Honkamaa 
Opponent: Associate Professor Mitko Veta, Eindhoven University of Technology, Netherlands
Custos: Professor Pekka Marttinen, Aalto University School of Science

Medical imaging is a fundamental part of modern clinical practice, enabling the capture of wide range of anatomical and physiological properties of imaged tissue. Different imaging techniques, such as magnetic resonance imaging (MRI) and computed tomography (CT), typically capture complementary information about tissue. For example, MRI can provide accurate information about tumor location, whereas CT can be used to estimate radiation absorption. When combining information from multiple images, mapping corresponding anatomical locations across images is often necessary to make use of this complementary information. This mapping or alignment process is called image registration. 

Despite extensive research, image registration remains challenging in many settings. Internal organs may undergo significant motion between image acquisitions, and patient posture can vary considerably. In addition, different imaging modalities highlight different aspects of tissue, making alignment difficult. Modern machine learning techniques have demonstrated superior performance in many medical image analysis tasks, and in recent years, substantial research has been conducted on medical image registration using so called neural networks. The hope is that neural-network-based methods could provide both fast and accurate registration. However, although these methods have shown promising results, neural networks have not yet demonstrated the same level of success in image registration as in many other tasks, highlighting the need for methodological innovations. 

This research proposes novel methods for solving the medical image alignment problem, with all proposed approaches employing neural networks. As an overarching theme, the thesis leverages the underlying mathematical structure of the problems to propose improved machine learning models. The developed methodologies have direct applications in clinical settings such as radiation therapy, enabling more accurate tumor targeting. On the methodological front, the ideas proposed in this work open up interesting new research opportunities that could lead to even better results in the future.

Keywords: medical image registration, deep learning, cross-modality image registration, cross-modality image synthesis, virtual staining, image-to-image translation

Thesis available for public display 7 days prior to the defence at Aalto University's public display page. 

Contact Information:
joel.honkamaa@aalto.fi