PASCAL - Pattern Analysis, Statistical Modelling and Computational Learning

Vocal Folds Analysis using Global Energy Tracking
Gal Elidan
Journal of Voice 2011.


Summary: Introduction. Detection and quantification of oscillatory irregularities in laryngeal videostroboscopy can be particularly difficult for the human expert. Accordingly, there is a wide interest in automated methods for recov- ering the folds’ temporal trajectory. Unfortunately, current methods typically provide only crude glottal measurements. Objectives. An automated procedure for consistently tracking the entire vocal folds’ boundary in laryngeal strobo- scopy videos, even when the glottal opening is closed. Methods. A preprocessing frame-by-frame crude mid point identification is followed by an active contour evolution to detect the global boundary in each frame independently. A global energy active contour is then jointly defined over the entire video sequence, and the full glottal boundary is detected throughout the video via standard energy minimization. Results. The vocal folds’ boundary is accurately tracked in normal and abnormal stroboscopy videos collected in a clinical setting, and that exhibit a varied range of visual characteristics (eg, lighting conditions). A proof-of-concept eval- uation based on the analysis of the waveform of the location of points along the boundary separates between a normal and two markedly different abnormal subjects, and automatically provides a hypothesized localization of the abnormality. Conclusion. The first method for automatically tracing the temporal trajectory of all points along the vocal folds’ boundary in all frames of a stroboscopy video is presented. The approach opens the door for novel analysis of all regions of the contour, which in turn may lead to automated localization of pathologies. Key Words: Vocal fold analysis–Automated tracking–Stroboscopy video–Active contours.

EPrint Type:Article
Project Keyword:Project Keyword UNSPECIFIED
Subjects:Machine Vision
Learning/Statistics & Optimisation
ID Code:9211
Deposited By:Gal Elidan
Deposited On:21 February 2012