Work

Deep Learning Approaches for Automatic Airway Invasion Detection and Kinematic Analysis on Videofluoroscopy

Year: 2023
Type: article
Abstract: Dysphagia is a significant health concern that affects millions of Americans per year. Identifying swallowing abnormalities and analyzing swallowing kinematics are crucial to early diagnosis and management of patients with suspected dysphagia. The videofluoroscopic swallow study is the criterion standard method for swallowing examination as it provides meticulous and dynamic visualization of swallowing anatomy and physiology. The interpretation of videofluoroscopy in clinical settings usually depends on clinicians' subjective judgment, which may be highly variable across different raters and diverse evaluation protocols. In addition, temporal and spatial kinematic measurements can be extracted to assess patient's swallowing function and potentially determine the pathophysiological cause of swallowing abnormality. Our first objective was to explore the association between airway invasion and temporal measurements of swallowing kinematics to add temporal and sequential understanding in clinical estimation of the risk of aspiration or penetration. However, these measurements on swallowing kinematics, that involve laborious and time-consuming frame-by-frame analysis, are rarely performed by clinicians. To gather efficient and reliable videofluoroscopic interpretations for swallowing diagnosis, we plan to apply advanced deep learning techniques to solve the following tasks. We aimed to develop a tool to automatically detect the top edge of tracheal column on videofluoroscopic images as an essential anatomical landmark for laryngeal elevation evaluation. Then we develop an automatic airway invasion detection method based on the laryngeal landmark positions. It was done by using a multitask learning technique that intentionally transfer laryngeal-related feature representations to the airway invasion prediction. The accuracy and interpretability of above landmark and airway invasion methods were examined. In addition, we investigated deep learning algorithms as data augmentation techniques in improving non-invasive aspiration detection performance using sensory swallowing screening tools. Overall, we hope to propose advanced solutions in performing automatic swallowing analysis on different modalities of swallowing examination to support swallowing diagnosis and prognosis in both clinical practice and research.
Author: Shu, Kechen
Language: English
Open Access status: green
OpenAlex ID: W7037235848
FWCI:
Cites:
Cited by:
Related to: 10