Unsupervised X-Ray-to-DRR Translation with Structural Enhancement for Efficient X-Ray-to-CT Registration

Artifacts and geometric distortions arising from modality differences are major sources of mismatch between conventional CT-based digitally reconstructed radiographs (DRRs) and real X-ray images. We propose an unsupervised crossmodal image translation framework that maps X-ray images to DRRs, enhancing structural features in the frequency domain while constraining geometric deformations in the spatial domain. Specifically, an edge-based spatial branch is introduced to guide the generator with richer structural information, while a Discrete Wavelet Transform (DWT)-based frequency enhancement module separates high-frequency details from low-frequency components to better preserve anatomical structures and improve cross-domain translation. Experiments on a public cadaveric pelvis dataset demonstrate state-of-the-art translation performance across multiple metrics. Moreover, when integrated into several advanced 2D-3D registration pipelines, the proposed method substantially improves registration accuracy and sub-millimeter success rates. The code is available at https://github.com/pj-jiang/unsupervised-x-ray-to-drr-images-translation.