Automated Dental Panoramic Image Segmentation Using Transfer Learning Based CNNs

Abstract

Over the years, deep learning technology improved considerably, and its application areas have expanded. At the same time, the content and size of the datasets used for deep learning have also increased. However, this is not the case for dental datasets. Therefore, in this paper the transfer learning method was used to overcome this disadvantage. Segmentation was performed on 131 dental panoramic X-ray images with two different models based on the transfer learning method. The first model was constructed with pre-trained U-Net using the chest X-ray dataset. As the second model, the pre-trained Inception-ResNet-v2 structure was used. The performances of the methods we developed were compared visually and quantitatively using dice coefficient, accuracy, and intersection over union. While the dice coefficient success of the first model was 87.12%, the success of the second model reached 90.26%. Our new approach of using transfer learning for dental image segmentation proved to be very successful.