Intraoperative 2D/3D Image Registration via Differentiable X-Ray Rendering

O p ti m iz a ti o n tr a je c to ry in (3 ) Pose regressor pertained on random poses over SE(3) Intraoperative optimization with differentiable rendering in (3) Fast registration of X-rays with sub-millimeter accuracy Initial guess t = 0.75 s t = 1.5 s Error = 14.4 mm 5.2 mm 0.53 mm Estimated pose Difference D is tr ib u ti o n o v e r ( 3 ) Synthetic X-rays Preop CT Encoder Predicted poses Encoder Intraop X-ray Figure 1. We present DiffPose, a self-supervised framework for differentiable 2D/3D registration. Trained exclusively on synthetic X-rays rendered from a patient-specific preoperative CT scan, DiffPose aligns intraoperative X-rays with sub-millimeter accuracy. DiffPose does not require manually annotated training data, performs consistently across subjects, and registers images at clinically relevant speeds.