Unsupervised learning for deformable medical image registration

Abstract

This thesis focuses on advancing unsupervised learning techniques for deformable medical image registration tasks. The primary objective is to develop generic and effective methods that address the challenges of deformable image registration. Three key contributions are presented: large deformation field decomposition, stochastic decomposition, and conformal invariant regularization. These methods are extensively evaluated using publicly available datasets, showcasing their superior performance compared to existing techniques and highlighting their potential for real clinical applications. The first contribution, large deformation field decomposition, tackles the complexity of large deformations in medical images by decomposing the deformation field into multiple continuous intermediate fields. A self-attention layer is employed to refine these intermediate fields, leading to enhanced accuracy in capturing complex and large deformations. The suggested method utilizes the temporal information presented in a breathing cycle, offering significant benefits for tasks like tumor tracking in image-guided systems. The second contribution, stochastic decomposition, introduces a novel training algorithm that effectively learns large deformation fields in medical image registration without the need of multiple networks or manual labels. This algorithm utilizes additional supervision information by stochastically decomposing the large deformation field and leveraging synthetic data with corresponding intensity discrepancies for registration output supervision. The results illustrate enhanced precision and robustness. The third contribution, conformal invariant regularization, presents a novel pair-wise image registration framework that eliminates the requirement for pre-training or prior affine registration. The framework incorporates a novel conformal invariant hyperelastic regularizer, which enforces the deformation field to be smooth, invertible and orientation-preserving. More importantly, the regularization strictly guarantees topology preservation yielding to a clinical meaningful registration. Additionally, a learned image deformation mapping is parameterized by coordinate MLP with periodic activation function, where one can view the to-be-registered images as continuously differentiable entities. In summary, this thesis addresses the challenges of large deformation learning in medical image registration tasks through novel methods and algorithms. It provides new perspectives and significant advancements in improving the accuracy and efficiency of deformable image registration.