Deep Learning for Video Fluoroscopic Swallowing Study Analysis: A Survey on Classification, Detection, and Segmentation Techniques

Deep learning has significantly advanced the analysis of Video Fluoroscopic Swallowing Study data, an essential diagnostic tool for dysphagia assessment. This review explores recent applications of deep learning across key VFSS analysis tasks, including classification, detection, and segmentation. Classification methods utilizing convolutional neural networks achieve high accuracy, ranging from 91.7% to 95.98%, and Area Under the ROC Curve scores between 0.71 and 0.97, thus enhancing the consistency and reliability of swallowing phase identification. Detection approaches employing advanced deep learning architectures effectively localize anatomical landmarks and temporal swallowing events, reaching Mean Average Precision values of up to 0.89 and tracking errors as low as 2.38 pixels. Segmentation techniques based on variants of U-Net and related architectures accurately delineate critical anatomical regions, with Dice Similarity Coefficients ranging from 0.67 to 0.90. Collectively, these advances substantially improve VFSS interpretation by increasing accuracy, reducing subjective variability, and streamlining clinical workflows. This survey summarizes recent methodologies and discusses strategies for dataset collection and preprocessing, including both proprietary and limited publicly available datasets, discusses ongoing challenges such as computational demands and dataset diversity, and highlights future directions in leveraging deep learning to enhance dysphagia diagnosis and treatment.