RESUMO
Anatomical segmentations generated using artificial intelligence (AI) have the potential to significantly improve video fluoroscopic swallow study (VFS) analysis. AI segments allow for various metrics to be determined without additional time constraints streamlining and creating new opportunities for analysis. While the opportunity is vast, it is important to understand the challenges and limitations of the underlying AI task. This work evaluates a bolus segmentation network. The first swallow of thin or liquid bolus from 80 unique patients were manually contoured from bolus first seen in the oral cavity to end of swallow motion. The data was split into a 75/25 training and validation set and a 4-fold cross validation was done. A U-Net architecture along with variations were tested with the dice coefficient as the loss function and overall performance metric. The average validation set resulted in a dice coefficient of 0.67. Additional analysis to characterize the variability of images and performance on sub intervals was conducted indicating high variability among the processes required for training the network. It was found that bolus in the oral cavity consistently degrades performance due to misclassification of teeth and unimportant residue. The dice coefficients dependence on structure size can have substantial effects on the reported value. This work shows the efficacy of bolus segmentation and identifies key areas that are detriments to the performance of the network.
