CTA-UNet: CNN-transformer architecture UNet for dental CBCT images segmentation.

Researchers

Journal

Modalities

Models

Abstract

Recent advances in deep learning are making it possible to segment dental Cone beam Computer Tomography (CBCT) images automatically and effectively. However, complex root morphological features, fuzzy boundaries between tooth roots and alveolar bone, and costly annotation of dental CBCT images limit the capability of existing deep learning models. To address the above issues, we collected dental CBCT data from 200 patients and labeled 45 of them for network training. We proposed the CNN-Transformer Architecture UNet (CTA-UNet) network, which can effectively extract local features by CNN and capture feature dependencies remotely by Transformer. Multiple spatial attention modules enhance the spatial information extraction and representation ability of the network. Further, we proposed a novel Masked image modeling (MIM) method to pre-train CNN and transformer modules simultaneously, to mitigate the limitations caused by a smaller amount of labeled training data. To the best of our knowledge, this is the first study that applied self-supervised learning methods to CNN-Transformer architecture network on dental CBCT images segmentation task. Experimental results show that the proposed method is superior to current dental CBCT image segmentation techniques and has real-world applicability in orthodontics and dental implants.Creative Commons Attribution license.