Self-Supervised Learning Methods for Label-Efficient Dental Caries Classification.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computational Dentistry

Background:

• High annotation costs hinder deep learning in clinical settings.
• Unlabeled data presents an opportunity for learning.
• Dental caries classification requires accurate diagnostic tools.

Purpose of the Study:

• To investigate the efficacy of self-supervised learning for dental caries classification.
• To reduce the reliance on extensively labeled datasets.
• To improve the label-efficiency of deep learning models in dentistry.

Main Methods:

• Trained three self-supervised algorithms on 38,000 unlabeled bitewing radiographs (BWRs).
• Applied learned representations to tooth-level dental caries classification using electronic health record (EHR) labels.
• Evaluated fine-tuned models on a holdout test set of 343 BWRs annotated by dental professionals.

Main Results:

• Pretraining with self-supervised algorithms improved caries classification performance.
• Achieved a 6 percentage point increase in sensitivity.
• Demonstrated significant label-efficiency, with models achieving >=45% sensitivity using only 18 annotations.

Conclusions:

• Self-supervised learning offers substantial gains in medical image analysis, especially for costly labeling tasks.
• This approach enhances diagnostic accuracy and reduces annotation burden in dental applications.
• The findings support the use of self-supervised methods for efficient and effective AI in healthcare.