Work

Detection of Vocal Fold Paralysis in Dynamic Translaryngeal Ultrasound Using a Deep Learning Approach

Year: 2025

Type: conference-paper

Abstract: This study introduces a new quantitative approach to assess vocal fold motion with dynamic Translaryngeal Ultrasound (dTU). Laryngeal cartilages have been annotated in dTU images during abduction or adduction phases, to train a YOLOv7-tiny model for detecting these cartilages. Motion features were extracted from object detection results and then used as inputs for a Support Vector Machine (SVM) classifier to diagnose vocal fold paralysis. A retrospective database was constructed consisting of 227 dTU scans, each with qualitative labels for normal and paralysed conditions determined by expert consensus. Experimental results showed that the YOLO model achieved a precision of 0.843 (resp. 0.644) and a recall of 0.827 (resp. 0.630) for the prominence of the thyroid (resp. arytenoid) cartilages on the test set of 46 scans. An univariate analysis of the motion features aligned with the expected physiological characteristics of healthy and paralysed vocal folds. Sensitivity and specificity values of SVM classifier with the YOLO-based method showed good results for the detection of true paralysis cases (sensitivity of 79%). The incorporation of additional labelled images is planned to further enhance classification accuracy.

Source: IEEE International Ultrasonics Symposium

DOI: 10.1109/ius62464.2025.11201284