Measurement of hyoid motion using submental ultrasound videos: reliability of manual measurements and validity of a semi-automated tracking method
10.3969/j.issn.1006-9771.2026.10.007
- VernacularTitle:颏下超声视频舌骨运动人工测量信度及半自动追踪的效度
- Author:
Zhenyuan NIU
1
;
Yaling YANG
1
;
Yuping DENG
1
;
Shanjia CHEN
1
;
Lei LEI
1
;
Xiaokuo HE
1
Author Information
1. Xiamen Fifth Hospital, Xiamen, Fujian 361100, China
- Publication Type:Journal Article
- Keywords:
hyoid motion;
swallowing;
submental ultrasonography;
semi-automated tracking;
deep learning;
reliability;
validity
- From:
Chinese Journal of Rehabilitation Theory and Practice
2026;32(10):1186-1193
- CountryChina
- Language:Chinese
-
Abstract:
ObjectiveTo establish a semi-automated workflow for hyoid motion tracking based on submental ultrasound videos, evaluate the reliability of manual measurements and the validity of semi-automated tracking, and compare the analytical efficiency. MethodsA total of 46 healthy adults were recruited at Xiamen Fifth Hospital from October, 2024 to August, 2025, and performed dry swallows and swallows of 5 mL and 10 mL moderately-thick liquid boluses. A total of 414 sagittal-plane submental ultrasound videos were acquired. After video quality screening and tracking quality-control procedures, 286 videos were included for paired analysis. A semi-automated tracking pipeline was constructed combining the channel-and-spatial-reliability tracker and You Only Look Once version 8 small (YOLO v8s). Intra-class correlation coefficients [ICC(A,1)] based on the absolute-agreement single-measurement model were calculated for test-retest and inter-rater manual measurements. Using the manual repeated-measurement results as the reference standard, Pearson correlation coefficient (r), ICC(A,1) and concordance correlation coefficient (CCC) were computed for the semi-automated measurements. A total of 5 000 subject-clustered bootstrap resamples were performed to calculate 95% confidence intervals. Cumulative analysis time of manual and semi-automated workflows was compared. ResultsFor the independent test set of YOLOv8s, precision and recall reached 0.940 and 0.956. The mean average precision was 0.974 at an intersection-over-union threshold of 0.5, and 0.609 averaged across thresholds ranging from 0.5 to 0.95. For manual measurements of maximum horizontal and vertical hyoid displacement, the test-retest ICC(A,1) values were 0.925 and 0.912, respectively; the inter-rater ICC(A,1) values were 0.815 and 0.775. For semi-automated versus manual reference measurements, r = 0.865，ICC(A,1) = 0.843，CCC = 0.843 in maximum horizontal displacement; and r = 0.754，ICC(A,1) = 0.746，CCC = 0.745 in maximum vertical displacement. Mean biases were -0.772 mm and -0.229 mm, with 95% limits of agreement of -4.567 to 3.023 mm and -2.918 to 2.460 mm for horizontal and vertical directions. Cumulative manual analysis time was 526 minutes for rater A and 713 minutes for rater B. The semi-automated workflow consumed 155 minutes, in which 40 minutes for computational processing and 115 minutes for manual review. ConclusionManual measurement of maximum hyoid displacement is reliable using qualified submental ultrasound videos. Semi-automated tracking achieves acceptable agreement with manual reference and greatly reduces analysis workload.
