Automated Penetration-aspiration Scale Scoring with Deep Learning

Research Objectives: To develop a deep learning algorithm that can automatically analyze Penetration-aspiration scale (PAS) and present its accuracy.

Design: Analysis of retrospectively collected data

Setting: Tertiary care hospital

Participants: A total of 1039 videofluoroscopic swallowing study (VFSS) video clips

Interventions: Two experienced rehabilitation physicians consensually assigned PAS scores for data coding. Regions of interest (ROI) were established in the oral and pharyngeal regions of the video and deep learning model was designed to automatically classify PAS. The structure of the model was to find ROIs using a convolutional neural network, which extracts visual features related to penetration and aspiration of ROIs and classifies them based on those features without any further manipulation of the video clips.

Main Outcome Measures: Predictive accuracy of the deep learning algorithm

Results: Out of 1039 video clips, the frequencies for PAS 1 through 8 were, in order, 226, 145, 178, 37, 137, 15, 132, 129. PAS 4 and 6 were excluded from training due to their low frequency, and 7 and 8 were combined as they were not discriminated in the videos. As a result, the classification accuracy was 76.09%, 42.31% 70.97%, 69.23%, and 75.00%, respectively, and the overall accuracy was 68.93%.

Conclusions: We have conducted research on using deep learning to automatically score PAS and have shown good results, though not high accuracy. In the future, we will need to train with more datasets to improve accuracy.

Author(s) Disclosures: We have no conflict of interest.