Baum, Zachary Michael Cieman; (2023) Medical Image Registration Using Deep Neural Networks. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text Baum_Thesis_v6-MinorRevision.pdf - Accepted Version Download (41MB) | Preview

Abstract

Registration is a fundamental problem in medical image analysis wherein images are transformed spatially to align corresponding anatomical structures in each image. Recently, the development of learning-based methods, which exploit deep neural networks and can outperform classical iterative methods, has received considerable interest from the research community. This interest is due in part to the substantially reduced computational requirements that learning-based methods have during inference, which makes them particularly well-suited to real-time registration applications. Despite these successes, learning-based methods can perform poorly when applied to images from different modalities where intensity characteristics can vary greatly, such as in magnetic resonance and ultrasound imaging. Moreover, registration performance is often demonstrated on well-curated datasets, closely matching the distribution of the training data. This makes it difficult to determine whether demonstrated performance accurately represents the generalization and robustness required for clinical use. This thesis presents learning-based methods which address the aforementioned difficulties by utilizing intuitive point-set-based representations, user interaction and meta-learning-based training strategies. Primarily, this is demonstrated with a focus on the non-rigid registration of 3D magnetic resonance imaging to sparse 2D transrectal ultrasound images to assist in the delivery of targeted prostate biopsies. While conventional systematic prostate biopsy methods can require many samples to be taken to confidently produce a diagnosis, tumor-targeted approaches have shown improved patient, diagnostic, and disease management outcomes with fewer samples. However, the available intraoperative transrectal ultrasound imaging alone is insufficient for accurate targeted guidance. As such, this exemplar application is used to illustrate the effectiveness of sparse, interactively-acquired ultrasound imaging for real-time, interventional registration. The presented methods are found to improve registration accuracy, relative to state-of-the-art, with substantially lower computation time and require a fraction of the data at inference. As a result, these methods are particularly attractive given their potential for real-time registration in interventional applications.

Download activity - last month

Export as CSVXMLJSON

Download activity - last 12 months

Export as JSONCSVXML

Downloads by country - last 12 months

Export as XMLCSVJSON

Archive Staff Only

View Item