Enhancing dental caries classification in CBCT images by using image processing and self-supervised learning.

Area of Science:

• Dentistry
• Artificial Intelligence
• Medical Imaging

Background:

• Diagnosing dental caries is challenging, requiring precise and early detection for effective management.
• Limited annotated medical images hinder the development of accurate diagnostic models.
• Cone Beam Computed Tomography (CBCT) provides detailed 3D imaging crucial for dental diagnostics.

Purpose of the Study:

• To improve dental caries classification in CBCT images using Self-Supervised Learning (SSL) tasks.
• To investigate the effectiveness of SSL in utilizing unlabeled data for enhanced model performance.
• To identify optimal image processing techniques, deep learning architectures, and SSL approaches for caries detection.

Main Methods:

• Developed a pipeline for unlabeled data extraction from CBCT exams.
• Employed various Self-Supervised Learning (SSL) tasks for model training.
• Integrated image processing techniques with SSL tasks.
• Evaluated different deep learning architectures (e.g., ResNet-18) and SSL approaches (e.g., SimCLR).

Main Results:

• ResNet-18 combined with the SimCLR SSL task achieved an average F1-score macro of 88.42%, Precision macro of 90.44%, and Sensitivity macro of 86.67%.
• SSL models showed a 5.5% increase in F1-score compared to models trained solely on deep learning architectures.
• The study explored the impact of unlabeled dataset sizes and the necessity of unlabeled data.

Conclusions:

• Self-Supervised Learning (SSL) significantly enhances the accuracy and efficiency of dental caries classification in CBCT images.
• SSL effectively addresses the challenge of limited annotated medical data in dentistry.
• The findings support the integration of SSL for improved diagnostic capabilities in dental imaging.