Automated detection of chewing movements in videofluoroscopic swallowing studies using deep learning for landmark detection and motion analysis

Chewing plays a critical role in safe and effective swallowing by breaking food into manageable pieces and mixing it with saliva to form a cohesive bolus. Impairments in this process, commonly seen in older adults and individuals with neurological conditions, increase the risk of choking, aspiration, and malnutrition. Although the videofluoroscopic swallowing study (VFSS) is considered the gold standard for evaluating swallowing function, automated analysis of chewing during VFSS remains largely unexplored. Manual assessments are time-consuming, variable, and rarely integrated into standard clinical protocols. This study presents the first fully automated pipeline for analyzing jaw movements related to chewing in VFSS. The pipeline integrates three main modules: (1) a landmark detection network to identify key anatomical points, (2) a video segmentation module to isolate the oral processing phase based on bolus progression, and (3) a classification module to distinguish chewing from non-chewing jaw movements. The system was trained and validated using datasets from multiple clinical studies involving both healthy participants and individuals with dysphagia. Results show high performance across all modules, demonstrating the feasibility and clinical relevance of the proposed approach. By enabling automated detection and classification of chewing cycles, this method offers a scalable tool for assessing masticatory function in clinical and research settings. It lays the foundation for future studies investigating how chewing efficiency impacts swallowing safety and may support the integration of mastication analysis into established VFSS protocols.