Self-supervised learning enhances periapical films segmentation with limited labeled data.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate segmentation of periapical dental films is crucial for diagnosis but relies on costly, large labeled datasets.
• Annotation variability in manual segmentation introduces inconsistencies and challenges for AI model development.

Purpose of the Study:

• To develop a self-supervised learning framework for periapical film segmentation that minimizes reliance on extensive labeled data.
• To enhance the practical applicability of AI in dentistry by reducing manual annotation efforts.

Main Methods:

• A two-stage framework involving self-supervised pre-training of a Vision Transformer (ViT) using DINOv2 weights on unlabeled periapical films.
• Further pre-training using student-teacher contrastive learning on 74,292 unlabeled periapical films.
• Fine-tuning the domain-adapted ViT with a Mask2Former head on only 229 labeled films for segmenting seven dental structures.

Main Results:

• The self-supervised method achieved a significantly higher Dice coefficient (74.77%) compared to traditional supervised models (33.53%-41.55%).
• The DINOv2-based approach outperformed other state-of-the-art self-supervised learning methods, including MAE, MoCov3, and BEiTv3.
• The method demonstrated statistically significant superiority over its supervised Mask2Former counterparts (p < 0.01).

Conclusions:

• The proposed two-stage, domain-specific self-supervised framework effectively learns robust anatomical features for accurate periapical film segmentation with minimal annotations.
• This approach addresses the challenge of limited labeled data in medical imaging and offers a feasible pathway for AI-assisted diagnostic tools.
• It promises to improve diagnostic accuracy and enhance workflow efficiency by reducing manual analysis time in dental practices.