Disclosures: Seong Jae Lee, MD, PhD: No financial relationships or conflicts of interest

Objective: Videofluoroscopic swallowing study (VFSS) is regarded as the standard tool for evaluation of the swallowing function. However, it is vulnerable to human errors and demands a high level of concentration and fatigue. Analysis of VFSS may be more accurate and efficient if human effort can be lessened with deep learning technologies. The aim of this study is to make a deep learning model that can track the hyoid bone, one of the most important landmarks in VFSS.

Design: Retrospective case control study Setting : University academic hospital Participants : A total of 31 VFSS video files were selected from the storage database. Each files were separated into 34,772 frame images and distributed for training, validation and test at the ratio of 7:1:2.

Interventions: Hyoid bone was labelled using the Computer Vision Annotation Tool (CVAT), with interpolation after human annotation every two frames. Training was performed by a new model in which weighted bi-directional FPN (BiFPN) was connected to Unet segmentation architecture for clear segmentation.

Main Outcome Measures: Overall accuracy, recall, precision, specificity

Results: Unet+BiFPN model showed overall accuracy of 99.9%, recall of 79.5%, precision of 87.1% and specificity of 99.9%. Average precision was 97.23%. Conclusions: Unet-BiFPN model has been attempted for the first time in this study, for identification of hyoid bone in VFSS images in a fully automated manner. Accuracy was increased significantly compared with preceding study that used obsolete deep learning model. Further researches for tracking of hyoid bone are to be carried out.

Level of Evidence: Level II

« Back to AAPM&R Annual Assembly 2020

PM&R Meeting Abstracts - https://pmrjabstracts.org/abstract/automated-tracking-of-hyoid-bone-in-videofluoroscopic-swallowing-study-deep-learning-model/