Tooth Localization with Coarse-to-Fine Auto-Encoders

Localization of teeth is a prerequisite task for most of the computerized methods for dental images such as medical diagnosis and human identification. Classical deep learning architectures like convolutional neural networks and auto-encoders seem to work well for tooth detection, however, it is non-trivial because of the large image size. In this study, a coarse-to-fine stacked auto-encoder architecture is presented for detection of teeth in dental panoramic images. The proposed architecture involves cascaded stacked auto-encoders where sizes of the input patches increase with the successive steps. Only the detected candidate tooth patches are fed into the successive layers, thus the irrelevant patches are eliminated. The proposed architecture decreases the cost of detection process while providing precise localization. The method is tested and validated on a dataset containing 206 dental panoramic images and the results are promising