SMTLNet: Domain Prior-Inspired Tooth Segmentation Based on Self-Supervised Manifold Transfer Learning.

Area of Science:

• Medical Imaging
• Computer Vision
• Digital Dentistry

Background:

• Accurate tooth identification in cone-beam computed tomography (CBCT) is vital for digital dentistry.
• Teeth present challenges like high interclass similarity and fuzzy boundaries.
• Limited labeled samples hinder existing segmentation methods due to time-consuming annotations.

Purpose of the Study:

• To propose a novel self-supervised manifold transfer learning network (SMTLNet) for improved tooth segmentation in CBCT images.
• To reduce reliance on extensively labeled datasets by leveraging unannotated data.
• To enhance segmentation accuracy and anatomical precision for clinical applications.

Main Methods:

• An object-oriented self-supervised pretraining approach was developed to extract image representations from unannotated CBCT data.
• A manifold optimization strategy was employed to regularize the segmentation model, improving class separation.
• A multiscale boundary constraint module was integrated to address blurred tooth boundaries and extract boundary-aware features.

Main Results:

• SMTLNet achieved state-of-the-art performance, with Dice Similarity Coefficients (DSCs) of 91.8% (100% data) and 89.08% (20% data).
• Jaccard Similarities (JSs) reached 86.71% (100% data) and 82.87% (20% data), demonstrating effectiveness with limited data.
• Anatomical precision was maintained with Hausdorff Distances (HDs) of 1.41 mm (high-resource) and 2.35 mm (low-resource).

Conclusions:

• The SMTLNet method significantly improves tooth segmentation accuracy in CBCT images, particularly under limited labeled data conditions.
• The network's ability to handle challenging cases like impacted teeth and crowded dentition highlights its clinical applicability.
• This approach offers a robust solution for advancing digital dentistry workflows through precise and efficient tooth segmentation.