Enhanced snake based segmentation of vocal folds

We present a system to segment the medial edges of the vocal folds from stroboscopic video. The system has two components. The first learns a color transformation that optimally discriminates, according to the Fisher linear criterion, between the trachea and vocal folds. Using this transformation, it is able to make a coarse segmentation of vocal fold boundaries. The second component uses an active contour formulation recently developed for the Insight Toolkit to refine detected contours. Rather than tune the internal energy of our active contours to bias for specific shapes, we optimize image energy so as to highlight boundaries of interest. This transformation of image energy simplifies the contour extraction process and suppresses noisy artifacts, which may confound standard implementations. We evaluate our system on stroboscopic video of sustained phonation. Our evaluation compares points on automatically extracted contours with manually supplied points at perceived vocal fold edges. Mean deviations for points located on the minor axes of the vocal folds averaged 2.2 pixels across all subjects, with a standard deviation of 3.6.