Abstract
<title>Abstract</title> <p> Oropharyngeal dysphagia affects nearly 45% of at-risk populations, including Parkinson’s disease, post-stroke, and head and neck cancer (HNC) patients. To support diagnosis and rehabilitation planning, mano-fluoroscopy, which combines high-resolution impedance manometry (HRIM) and videofluoroscopic swallow studies (VFSS), is increasingly being used. However, these examinations remain difficult to interpret and complex to execute. Fully automated cross-modal registration could help address these limitations, but remains challenging as manometric sensors are often occluded or leave the fluoroscopic field of view. We present GHOST, a geometry-driven framework that combines deep learning-based sensor detection with a point-cloud representation, coherent point drift registration, and the Hungarian assotiation algorithm to preserve sensor identities throughout the examination. The method was evaluated in 16 examinations of 12 patients with HNC using three detector backbones (YOLO11n, YOLO12n and YOLO26n) under <italic>high quality</italic> and emulated <italic>low quality</italic> imaging conditions. Detection performance was consistently high across all models, with <italic>Precision</italic> , <italic>Recall</italic> , and <italic>F1-score</italic> values greater than 98%. GHOST substantially improved sensor identification over a per-frame sequentially ordered baseline in both experimental settings and across all detector backbones. In the <italic>high quality</italic> setting, GHOST achieved the highest identification recall with YOLO26n (96.6%), and remained robust in the <italic>low quality</italic> setting. These results demonstrate that accurate and stable cross-modal registration can be achieved without manual intervention, providing a practical foundation for computer-assisted analysis of mano-fluoroscopy examinations. Such tools may reduce the interpretive burden and improve the overall usefulness of mano-fluoroscopies. </p>