Deformable registration based data augmentation for tooth segmentation

Abstract

Over the past few decades, deep learning in medical imaging has played an important role in medical image segmentation which helps accurate diagnosis and efficient treatment planning. Using deep learning in medical image segmentation requires a large amount of data to train the network, however, there is a problem that a large amount of data is not always available. We proposed a data augmentation algorithm using deformable registration to train the network with a limited amount of data for accurate molar tooth segmentation. First of all, using deformable registration, we generated the motion vector field (MVF), which reflects the relationship between two images. Re-scaling the MVF and applying it to the original image, we can generate various data to help the network to train in diverse cases. By applying the MVF method, the score calculated by the dice coefficient increased to 0.9148 from the original method of 0.8876. Thus, we have demonstrated the effectiveness of applying deformable registration and MVF in the data augmentation step to train deep learning segmentation models from a limited amount of data.