Masked Deep Embeddings of Patches for Teeth Segmentation in Cone Beam CT Images

In modern dentistry, teeth localization and segmentation from dental cone-beam CT (CBCT) images are crucial for improving dental diagnostics, treatment planning, and population-based studies on oral health. However, creating automated algorithms for teeth analysis is a challenging task due to the limited availability of accessible data for training. This work extends the self-supervised learning framework of the masked autoencoder (MAE) transformer. While the MAE loss measures the quality of reconstructed masked patches, we propose to instead measure the closeness of the predicted deep embeddings of masked patches to their originals. This yields a better generalization ability on a very limited number of CBCT images, as documented by our results on teeth segmentation of CBCT images. We call our approach DEMAE for deep embedding MAE. Our results show that masking-based unsupervised learning methods may, for the first time, provide convincing transfer learning improvements on CBCT images, increasing the overall accuracy over both MAE and prior self-supervised pre-training.