YOLO11-based detection of manometry sensors in video-fluoroscopy imaging for computer-aided multimodal assessment of swallowing

Purpose: Accurate assessment of swallowing function is essential in the diagnosis and monitoring of dysphagia following head and neck cancer (HNC). The simultaneous analysis of video-fluoroscopy swallow studies (VFSS) and high-resolution impedance manometry (HRIM) offers a more comprehensive evaluation, reducing subjectivity in VFSS and improving anatomical context of HRIM in HNC patients. Methods: The inherently low pharyngeal pressures in post-treatment HNC patients hinder the analysis of HRIM. As such, this study proposes a deep learning method for the automatic detection of HRIM sensors in VFSS using a YOLO11-based detector, aimed at enabling the automatic delineation of manometric regions. Detection performance was evaluated on 268 frames from 8 HNC patients using a leave-one-patient-out cross-validation approach. EigenCAM-based heatmaps were produced to analyze the model's attention patterns. Results: The model achieved 95.8% Precision, 97.4% Recall, 96.6% F1-score with minimal variation between folds. Under different noise levels and bolus-simulated obstructions, performance remained robust. Our method outperformed previous template-matching methods for manometric sensor detection in VFSS. EigenCAM visualizations confirmed consistent attention to catheter regions. Conclusion: The proposed YOLO11-based detector provides accurate and robust localization of manometric sensors in VFSS sequences to facilitate computer-assisted HRIM-VFSS fusion for objective swallowing assessment.