UCLA

Deformable image registration (DIR) is an important component of a patient’s radiation therapy treatment. During the planning stage it combines complementary information from different imaging modalities and time points. During treatment, it aligns the patient to a reproducible position for accurate dose delivery. As the treatment progresses, it can inform clinicians of important changes in anatomy which trigger plan adjustment. And finally, after the treatment is complete, registering images at subsequent time points can help to monitor the patient’s health. The body’s natural non-rigid motion makes DIR a complex challenge. Recently neural networks have shown impressive improvements in image processing and have been leveraged for DIR tasks. This thesis is a compilation of neural network-based approaches addressing lingering issues in medical DIR, namely 1) multi-modality registration, 2) registration with different scan extents, and 3) modeling large motion in registration. For the first task we employed a cycle consistent generative adversarial network to translate images in the MRI domain to the CT domain, such that the moving and target images were in a common domain. DIR could then proceed as a synthetically bridged mono-modality registration. The second task used advances in network-based inpainting to artificially extend images beyond their scan extent. The third task leveraged axial self-attention networks’ ability to learn long range interactions to predict the deformation in the presence of large motion. For all these studies we used images from the head and neck, which exhibit all of these challenges, although these results can be generalized to other parts of the anatomy.The results of our experiments yielded networks that showed significant improvements in multi-modal DIR relative to traditional methods. We also produced network which can successfully predict missing tissue and demonstrated a DIR workflow that is independent of scan length. Finally, we trained a network whose accuracy is a balance between large and small motion prediction, and which opens the door to non-convolution-based DIR. By leveraging the power of artificial intelligence, we demonstrate a new paradigm in deformable image registration. Neural networks learn new patterns and connections in imaging data which go beyond the hand-crafted features of traditional image processing. This thesis shows how each step of registration, from the image pre-processing to the registration itself, can benefit from this exciting and cutting-edge approach.