A Registration-Based Framework for Generating Aligned CT Images from Misaligned CBCT-CT Pairs

Motivation: As Cone Beam Computed Tomography (CBCT) becomes more prevalent in clinical applications, accurately registering Computed Tomography (CT) to CBCT images is critical, yet challenging due to CBCT's lower image quality and complex anatomical distortions. Goal(s): Generate aligned CT images via spatial transformations and guided cross attention within a custom module. Approach: Registration by Generation (RbG), a self-supervised framework that uses misaligned CT images as a reference guiding Deformation-Aware Cross Attention (DACA), aligning CT to CBCT with high fidelity and structural consistency. Results: The proposed RbG method demonstrates superior performance compared to traditional and deep learning registration and image translation methodologies across all metrics. Impact: This methodology eliminates the need for perfectly aligned inputs, which are often unavailable in practice. By using the misaligned CT image as a proxy label, the proposed self-supervised approach leverages CBCT information to enhance the high-quality aligned CT (aCT) format.