Assessing the Performance and Reliability of Deep Learning Auto-Segmentation in Videofluoroscopic Swallowing Studies: A Systematic Review and Meta-Analysis.

Abstract

To systematically evaluate the accuracy and reliability of deep learning-based auto-segmentation methods in videofluoroscopic swallowing studies (VFSS) through meta-analysis. A comprehensive literature search was conducted across PubMed, IEEE Xplore, Embase, Web of Science, and Cochrane Library databases for studies published in English between 2013 and 2024. Studies were included if they applied deep learning techniques to the auto-segmentation of anatomical structures in VFSS, specifically the bolus, cervical spine, hyoid bone, or thyroid cartilage-vocal fold complex (TVC) and reported quantitative performance metrics such as the Dice similarity coefficient. Two independent reviewers extracted data on study characteristics, segmentation targets, deep learning model types, and performance metrics. Methodological quality was assessed using the CLAIM and QUADAS-2 tools. Ten studies met inclusion criteria. A random-effects meta-analysis yielded an overall pooled Dice score of 0.83 (95% CI: 0.76-0.88, I² = 77%). Subgroup analyses showed similar performance for bolus segmentation (pooled Dice score = 0.84; 95% CI: 0.70-0.92, I² = 74%) and cervical spine segmentation (pooled Dice score = 0.83; 95% CI: 0.69-0.91, I² = 87%). Despite high accuracy, substantial heterogeneity was observed. Deep learning-based auto-segmentation in VFSS demonstrates promising accuracy across different anatomical targets. However, methodological variability among studies underscores the need for standardized protocols, multi-center datasets, and comparative evaluations of model architectures to enhance generalizability and clinical utility. PROSPERO registration: CRD42024578117.