Detecting Airway Invasion in Variable-Length Videofluoroscopic Swallowing Studies: A Vision Transformer Approach for...

Background: Dysphagia from aging, neurodegeneration, structural anomalies, or cognitive decline harms quality of life. The videofluoroscopic swallowing study (VFSS) is the diagnostic gold standard but manual interpretation is labor-intensive and costly, motivating automation. Methods: We introduce a Vision Transformer (ViT) using a temporal sliding window and 3D patch tokenization to capture spatio-temporal dependencies in variable-length VFSS via attention. Training/evaluation used 1154 VFSS sequences from 107 individuals (548 abnormal, 606 normal) with 5-fold cross-validation and comparisons to VGG-16, ResNet-50, EfficientNet-V1/V2, and MobileNet. Results: The ViT achieved 84.37 ± 1.15% accuracy, 90.81 ± 2.11% sensitivity, 79.49 ± 1.66% specificity, 82.94 ± 2.76% precision, 85.68 ± 1.54% F1-score, and AUC 0.878 (5-fold). It outperformed all CNN baselines across metrics; paired t-tests confirmed significant gains (p < 0.05). Conclusions: The pure ViT's attention-based spatio-temporal modeling yields robust VFSS classification and is well-suited for screening workflows requiring timely abnormality detection, providing a foundation for clinically deployable VFSS analysis.