A multiscale product approach for an automatic classification of voice disorders from endoscopic high-speed videos

Direct observation of vocal fold vibration is indispensable for a clinical diagnosis of voice disorders. Among current imaging techniques, high-speed videoendoscopy constitutes a state-of-the-art method capturing several thousand frames per second of the vocal folds during phonation. Recently, a method for extracting descriptive features from phonovibrograms, a two-dimensional image containing the spatio-temporal pattern of vocal fold dynamics, was presented. The derived features are closely related to a clinically established protocol for functional assessment of pathologic voices. The discriminative power of these features for different pathologic findings and configurations has not been assessed yet. In the current study, a collective of 220 subjects is considered for two- and multi-class problems of healthy and pathologic findings. The performance of the proposed feature set is compared to conventional feature reduction routines and was found to clearly outperform these. As such, the proposed procedure shows great potential for diagnostical issues of vocal fold disorders.