Automatic glottal segmentation using local-based active contours

International audienceHigh-speed videoendoscopy is the most promising approach to directly assess vocal-fold vibrations. Yet, its application to clinics is limited by the vast amount of data that has to be evaluated both qualitatively and quantitatively. There is need to reduce the dimensionality of the spatio-temporal information, and to efficiently represent the high-speed data in a compact, handy and lossless way. In this issue, automatic segmentation of glottal area is a major challenge. Recently, several advanced methods have been proposed for glottal segmentation, based either on a region-growing approach (Yan et al., 2006; Lohscheller et al., 2007; Demeyer et al., 2009) or on an active-contours framework (Marendic et al., 2001; Allin et al., 2004; Moukalled et al., 2009). The choice of segmentation method is not trivial, as it depends on image quality, the features profile of the object of interest, and computational demands. In this paper, we wish to explore the possibilities and limitations of an active-contours based method for automatic segmentation with no user intervention. The basic principle of active-contours models, also know as snakes, is to consider the object detection as a problem of energy minimization. An energy-minimizing spline is guided by external constraint forces and influenced by image forces, that pull it towards desired features such as lines and edges (Kass et al., 1988). The dynamic glottal-edge detection applied here is based on the local region-based framework proposed by Lankton and Tannenbaum (2008). It allows the foreground and background to be modeled in terms of smaller local regions, instead of representing them with global statistics. The main contributions of the proposed approach are the following. First, a precise way of localizing the glottal area is presented. The proposed scheme provides a tight mask surrounding the object of interest, information on the maximum area threshold and information on the presence or absence of glottal area. Second, local-based active contours are used to resolve convergence failure in global energy for objects with heterogeneous statistics. The method allows the contour to split and merge, thus dealing efficiently with cases where the glottal area consists of more than one regions, as is often the case in pathological vocal-fold vibrations. Furthermore, the method is fully automatic and it does not require human intervention. Parameter selection is performed automatically, based on the sequence statistics and empirical observations performed on a wide range of high-speed sequences. Finally, the proposed method is not data-dependent. The method has been tested on a database of 60 high-speed sequences of 501 frames (resolution of 256*256 interlaced pixels and duration equal to 125 msec), recorded from two subjects performing several speech and singing tasks (mainly glides, and sustained phonation with different voice qualities). The high-speed images were recorded at a frame rate of 4000 fps, simultaneously to audio and electroglottographic signals. The 60 sequences were verified by visual inspection and manually corrected when needed. Phonovibrograms were computed for qualitative validation (Lohscheller et al., 2008). Manual verification resulted on average in +/- 10 pixels modifications, which corresponds to less than 1% of the average glottal area. These errors mainly come from detection failure in the posterior or anterior parts of the glottal area. In 38/60 sequences, the amount of mean error relative to glottal opening and closing instants was less than 2 pixels, which validates the use of the segmentation results for glottal features estimation. A comparison has been conducted on glottal-parameters (fundamental frequency , open quotient) estimated either on electroglottographic signal or on the extracted glottal area (de Cheveigné and Kawahara, 2002; Henrich et al., 2004, respectively), using the electroglottographic signal as reference. Only 5/60 sequences exhibited mean differences in f0 greater than 0.5 semitones. A good match was found for open quotient in half of the sequences. References Allin, S., Galeotti, J., Stetten, G., Dailey, S.H. (2004) Enhanced snake based segmentation of vocal folds, IEEE International Symposium on Biomedical Imaging: Nano to Macro, 812- 815 Demeyer, J., Dubuisson, T., Gosselin, B., Remacle, M. (2009) Glottis segmentation with a high-speed glottography : a fully automatic method, 3rd Advanced Voice Function Function Assessment International Workshop. De Cheveigné, A., Kawahara, H. (2002) YIN, a fundamental frequency estimator for speech and music, The Journal of the Acoustical Society of America, vol. 111, 1917-1930,2002 Henrich, N., d'Alessandro, C., Doval, .B, Castellengo, M. (2004) On the use of electroglottographic signals for characterization of nonpathological phonation, The Journal of the Acoustical Society of America, vol.115, no. 3, 1321-1332. Kass, M., Witkin, A., Terzopoulos, D. (1988) Snakes: Active contour models International journal of computer vision, vol. 1, no. 4, 321-331. Lankton, S., Tannenbaum, A. (2008) Localizing region-based active contours IEEE Transactions on Image Processing, vol. 17, no. 11, 1-11. Lohscheller, J., Toy, H., Rosanowski, F., Eysholdt, U., Dollinger, M. (2007) Clinically evaluated procedure for the reconstruction of vocal fold vibrations from endoscopic digital high-speed videos Medical Image Analysis, vol. 11, no. 4, 400-413 Lohscheller, J., Eysholdt, U., Toy, H., Dollinger, M. (2008) Phonovibrography: Mapping high-speed movies of vocal fold vibrations into 2-D diagrams for visualizing and analyzing the underlying laryngeal dynamics IEEE trans- actions on medical imaging vol. 27, no. 3, 300-309. Marendic, B., Galatsanos, N., Bless, D. (2001) New active contour algorithm for tracking vibrating vocal folds International Conference on Image Processing, Proceedings, vol. 1. Moukalled, H. J., Deliyski, D. D, Schwarz, R. R, Wang, S. (2009) Segmentation of laryngeal high-speed videoendoscopy in temporal domain using paired active contours MAVEBA. Yan, Y., Chen, X., Bless, D. (2006) Automatic tracing of vocal-fold motion from high-speed digital images. IEEE transactions on bio-medical engineering, vol. 53, no. 7, 139