Unsupervised Keypoint-Based 2D/3D Registration in Minimally Invasive Pelvic Surgery

2D/3D medical image registration plays a critical role in pelvic surgeries by aligning preoperative 3D CT scans with intraoperative 2D X-ray images. Most existing methods simplify this task into <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$2 \mathrm{D} / 2 \mathrm{D}$</tex> registration using digitally re-constructed radiographs (DRRs). However, optimization-based approaches are often computationally expensive, sensitive to initial pose estimates, and lack real-time capability. Although recent learning-based methods improve speed and accuracy by directly regressing transformation parameters, they typically require patient-specific training and fail to fully exploit 3D spatial information, limiting their generalizability. To overcome these limitations, we propose an unsupervised keypoint-based 2D/3D pelvic registration framework. A GAN-based sparse-view X-ray-to-CT reconstruction network first synthesizes pseudo-3D CT volumes from 2D X-rays. These are then paired with preoperative CT scans and fed into an unsupervised registration network, which automatically identifies anatomical keypoints and estimates spatial transformations by minimizing Euclidean distances via closed-form optimization. Experimental results demonstrate that the proposed method achieves superior registration accuracy and computational efficiency compared to existing approaches, highlighting its strong potential for clinical application in real-time surgical navigation.