RGB-D/CT Registration for Preemptive Patient Pose Alignment in Longitudinal CT Scans

Abstract

Registration of longitudinal Computed Tomography (CT) scans is critical for monitoring treatment responses, yet its precision is often compromised by inconsistencies in patient positioning between sessions. Historically, the medical imaging community has relied purely on post-hoc ("postscan") image registration to correct these inconsistencies [1]. Whether utilizing rigid or deformable models, post-hoc registration is a reactive tool applied to data that is inherently flawed at acquisition. This work proposes a paradigm shift: addressing variability at its source through pre-scan patient pose alignment. By capturing the patient’s current pose with an integrated RGB-D camera prior to the follow-up CT scan, discrepancies relative to a baseline CT scan can be registered and visualized. This paper outlines an articulated registration pipeline designed to overcome the domain gap between occluded RGB-D surface data and volumetric CT data. The torso serves as a rigid anchor to align the modalities, after which individual limbs undergo local point-to-plane Iterative Closest Point refinement. The spatial corrections are then projected onto the original 2D RGB image, providing intuitive visual guidance for clinical staff. Validated on an anthropomorphic phantom, our pipeline demonstrates highly stable torso anchoring (average centroid error over 10 iterations: 2.69 cm) and consistent limb alignment (mean surface error: 1.61-1.84 cm). This proof-of-concept establishes a foundation for proactive patient alignment in longitudinal imaging.