Semantic Segmentation Method for Tooth Type with Data Augmentation Using Alpha-Blending

The purpose of this study was to develop a semantic segmentation method for each tooth type by DeepLab v3+ with data augmentation using alpha-blending. Our database consisted of 119 dental panoramic X-ray images and 119 truth label images for 32 tooth types. In the proposed method, training dataset was augmented using alpha-blending to be more focused to tooth shape in training DeepLab v3+. Mean intersection over union (IoU) for each tooth with the proposed method was 0.671, which was greater than that for DeepLab v3+ without alpha-blending (0.643).