Artificial Intelligence for Detection of External Cervical Resorption Using Label-Efficient Self-Supervised Learning...

Area of Science:

• Artificial Intelligence in Dentistry
• Medical Imaging Analysis
• Machine Learning for Diagnostics

Background:

• Detecting endodontic-periodontal lesions (ECR) and differentiating them from caries is crucial for accurate dental diagnosis.
• Traditional methods often require large, labeled datasets, which can be time-consuming and expensive to acquire.

Purpose of the Study:

• To evaluate the efficacy of label-efficient self-supervised learning (SSL) models in detecting ECR and distinguishing it from caries.
• To compare the performance of various SSL models against traditional deep learning baselines for radiographic analysis.

Main Methods:

• Collected periapical (PA) radiographs of teeth with and without ECR defects.
• Utilized two endodontists for ground truth determination using PA and CBCT imaging.
• Implemented and assessed nine SSL models (e.g., DINO, MoCo v2) and seven baseline deep learning models.
• Employed 10-fold cross-validation and a hold-out test set for robust model evaluation.

Main Results:

• Most SSL models significantly outperformed transfer learning baselines.
• The DINO model achieved the highest mean accuracy (85.64%) and test set accuracy (84.09%).
• MoCo v2 demonstrated the highest recall (77.37%) and F1-score (82.93%).

Conclusions:

• AI, particularly SSL-based models, can effectively assist clinicians in detecting ECR and differentiating it from caries.
• SSL models offer a viable alternative to traditional methods, reducing the dependency on large, labeled datasets.
• This study highlights the potential of SSL in advancing dental diagnostics through improved radiographic analysis.