New Deep Learning-Based Approach for Dysphagia Assessment from Videofluoroscopy Swallowing Studies

Dysphagia, one of the main toxicities from head and neck (HN) cancer radiotherapy, is usually assessed through Videofluoroscopy swallowing studies (VFSS). However, little effort has been directed to the automatic assessment of VFSS. In this study, we propose a deep learning-based iterative framework to detect ten regions of interest (ROIs) for dysphagia evaluation from VFSS. Moreover, we present a novel approach for parameter analysis from the automatically labeled videos. The database used includes 15 HN cancer patients with dysphagia with VFSS acquired before and after radiotherapy treatment, and 15 healthy participants as controls. Results show that the labeling framework detects the ten ROIs with an average distance error smaller than 2 pixels for the test set and 2.5 pixels for the external cohort. Furthermore, the characterization of several dynamic parameters yields statistically significant differences between groups. These findings support the robustness of our proposal to characterize dysphagia from VFSS objectively.