NKUT: Dataset and Benchmark for Pediatric Mandibular Wisdom Teeth Segmentation.

Area of Science:

• Dentistry
• Medical Imaging
• Artificial Intelligence

Background:

• Germectomy for impacted mandibular wisdom teeth is common in pediatric dentistry.
• Manual segmentation of teeth and bones from 3D volumes is time-consuming and delays treatment.
• Deep learning requires extensive annotated data for medical image segmentation.

Purpose of the Study:

• To curate the first publicly available Cone Beam Computed Tomography (CBCT) dataset for pediatric mandibular wisdom teeth segmentation.
• To propose a novel deep learning network (WTNet) for accurate teeth and bone segmentation.
• To address feature confusion and semantic blur in medical image segmentation tasks.

Main Methods:

• Curated the NKUT dataset, the first public CBCT dataset for pediatric mandibular wisdom teeth segmentation.
• Developed WTNet, a semantic separation scale-specific feature fusion network with two branches for teeth and bone segmentation.
• Designed Input Enhancement (IE) and Teeth-Bones Feature Separation (TBFS) blocks within WTNet to resolve feature confusions and semantic blur.

Main Results:

• WTNet demonstrated superior performance on the NKUT dataset compared to state-of-the-art methods like TransUnet.
• Achieved a maximum Dice Similarity Coefficient (DSC) improvement of nearly 16% over existing methods.
• The proposed IE and TBFS blocks effectively addressed feature confusion and semantic blur.

Conclusions:

• The NKUT dataset is a valuable resource for advancing pediatric mandibular wisdom teeth segmentation research.
• WTNet offers a significant improvement in automated segmentation accuracy for dental CBCT images.
• This work has the potential to reduce manual annotation burden and expedite surgical planning in pediatric dentistry.