Model-driven unsupervised medical image registration

Jia, Xi (2024). Model-driven unsupervised medical image registration. University of Birmingham. Ph.D.

Preview

Jia2024PhD.pdf
Text - Accepted Version
Available under License All rights reserved.
Download (19MB) | Preview

Abstract

In this thesis, we study pairwise medical image registration, a process that involves determining the optimal spatial deformation between two medical images. While traditional image registration methods relied on iterative optimization, recent years have witnessed a significant shift towards modern deep data-driven techniques. Within this context, this thesis focuses on unsupervised neural network-based registration approaches, with a primary emphasis on integrating prior knowledge derived from model-based strategies into network architectures to enhance their learning capabilities. More specifically, we address three critical research challenges currently facing the medical image registration community. The first challenge centers on whether deep data-driven registration methods can achieve higher registration accuracy. The second challenge concerns enhancing the data efficiency of deep registration methods. The third challenge involves accelerating the speed of deep registration methods, in terms of both training and inference.

To address the first challenge, we begin by offering a comprehensive guide to constructing a basic deformable registration method based on the U-Net architecture. Alongside this fundamental model, we then introduce three novel variants: Large Kernel U-Net (LKU-Net), which increases the network’s effective receptive field through multiple parallel convolutional layers to capture finer spatial structures from the input data; Cascaded LKU-Net, which stacks multiple base LKU-Nets and thereby effectively manages large deformations; and LKU-Net-Affine, which incorporates multiple fully connected modules to learn transformation-specific parameters, enabling adaptability to parametric rigid and affine registration tasks. These variants empower the U-Net model to address both parametric and deformable registration and effectively capture complex deformations that may present in high-resolution medical image data, such as 3D expansion microscopy images and 3D lung CT images.

To address the second challenge, we introduce VR-Net, a variational registration network that unrolls the mathematical structure of iterative variational optimization through variable splitting and seamlessly integrates it into a deep neural network in a cascading fashion. VR-Net is designed to inherit prior knowledge drawn from model-based solvers, ensuring the preservation of their data efficiency and generalizability while maintaining the fast speeds of learning-based approaches. Extensive experiments conducted on 2D and 3D cardiac MRI datasets demonstrate that VR-Net outperforms both traditional model-based and deep data-driven approaches in terms of registration accuracy while retaining fast inference speed and data efficiency.

To address the third challenge, we present Fourier-Net, a novel approach that predicts a compact, low-dimensional representation of the deformation in the band-limited Fourier space. Within Fourier-Net, we propose a model-driven decoder to effectively reconstruct the full-resolution deformation field from the band-limited coefficients. We then introduce Fourier-Net+, an extension of Fourier-Net, which learns the band-limited deformation field from the band-limited representation of images. Fourier-Net+ further accelerates the registration speed by constraining both the input and output of the network to low-dimensional representations, reducing the need for repeated convolution operations. Fourier-Net and Fourier-Net+ are evaluated on cardiac and brain MRI registration tasks in both 2D and 3D scenarios, demonstrating significant reductions in multiply-addition operations, memory footprint, and CPU runtime.

Type of Work:

Thesis (Doctorates > Ph.D.)

Award Type:

Doctorates > Ph.D.

Supervisor(s):