Preprocessing Optimization and Semantic Segmentation for Extraction of Cervical Intervertebral Disks from Videofluorography

Semantic segmentation (SS) is one of the most powerful tools to extract particular regions from medical images. Several studies succeeded to increase the segmentation accuracy of SS by combining with preprocessing. The present study proposed an extraction method of cervical intervertebral disks from videofluorography (VF), which is commonly used for the diagnosis of dysphagia, based on SS combined with 45 linear and nonlinear image filters. The combination of the image filters was optimized by the simulated annealing algorithm. In this study, five fully convolutional networks (FCNs), i.e. U-Net, feature pyramid network, LinkNet, pyramid scene parsing network and M-Net, were applied to the VF dataset of 19 patients and 39 healthy participants. When the image filters were not used, the mean F measures of the five FCNs were 0.660, 0.752, 0.803, 0.750 and 0.768, respectively, whereas when used, they were increased to 0.747, 0.794, 0.813, 0.765 and 0.799, respectively. This experimental results demonstrated that the optimal combinations of image filters were effective to improve the segmentation accuracy of FCNs.