Tracking of Bodies of Hyoid Bones in Videofluorography by Use of Support Vector Machine

Abstract

Swallowing phases are defined on the basis of the positions and moving directions of bodies of hyoid bones (BHBs) in videofluorography (VF). In this report, we propose an automatic tracking method of BHBs by use of Support Vector Machine (SVM). First, the movable prediction areas of BHBs are determined by applying discriminant analysis, morphology operations, and Random Sample Consensus techniques to frame images in VF. Then, a patchbased SVM is applied to the movable prediction areas to extract the candidate regions of BHBs. In general, several candidate regions are extracted in each frame. The false positives are eliminated by considering the continuity of the candidate regions between frame images. The proposed method was applied to actual twenty cases, and the tracking accuracy was evaluated on the basis of (1) distances between the extracted and truth regions, and (2) the correlation coefficients of x and y coordinates of these regions. The mean distance was approximately 2.9 mm, and the correlation coefficients were 0.906 and 0.958 in the x and y coordinates, respectively.