Deep learning-assisted creation of a large-scale cone beam computed tomography dataset for the segmentation of impacted mandibular third molar.

Background: Impacted third molar (ITM) extraction is a common procedure in oral and maxillofacial surgery. The segmentation of ITM is a crucial step in surgery planning. Manually segmenting ITM via cone beam computed tomography (CBCT) images requires significant time and effort from clinicians and may lead to treatment delays. With the advancement of deep learning, medical image segmentation is gradually becoming automated, but it still requires large amounts of well-annotated data for training. The aim of this study was to construct a large-scale CBCT dataset with the assistance of deep learning methods for the segmentation of ITM. Methods: A large-scale CBCT dataset for mandibular ITM segmentation, named "ITM-mandibular nerve (ITM-MN)", was constructed and comprised 1,262 CBCT image volumes, 186 expert manual annotations, and 1,076 deep learning-assisted annotations. The AttentionUnet3D model, based on an attention mechanism, was employed with a merged annotation strategy to adaptively focus on important feature regions for segmentation tasks. Results: On the ITM-MN dataset, AttentionUnet3D achieved the best segmentation results compared to other models, with an average Dice coefficient of 86.46%, an average recall of 84.69%, and a Hausdorff distance of 48.49 mm. The speed of deep learning-assisted annotation was only 16.51±3.78 seconds per case, with quality assessment scores of 3.07±0.66 and 3.29±0.74 from two physicians, respectively. Conclusions: ITM-MN is the first large-scale CBCT dataset specifically designed for segmenting ITM. The AttentionUnet3D model achieved effective automatic segmentation of ITM, second molars, and the MN.