Self-Supervised Learning for Dental Image Analysis

Objectives: A common bottleneck for supervised deep learning in medical and dental image analysis is the availability of expert-annotated data. In this study we assessed the value of self-supervised learning for dental image analysis.
Methods: ResNet-18, a common Convolutional Neural Network architecture, was trained using random initialization versus supervised pretraining on the ImageNet dataset versus self-supervised pretraining using a Momentum Contrast (MoCo) model. Pretraining was done on 3457 unlabeled bitewings radiographs. All models were fine-tuned on 386 expert-annotated radiographs (269 images with caries being segmented pixelwise, 116 without).
Results: Self-supervised pretraining led to significantly higher performance than random initialization (F1-score: 0.435 [0.352, 0.519] vs 0.279 [0.223, 0.336], Negative Predictive Value:0.669 [0.498, 0.839] vs 0.257 [0.016, 0.498], Area-Under-the-Curve: 0.546 [0.506, 0.586] vs 0.497 [0.493, 0.502]) (95% confidence intervals). In contrast, performance was not significantly different between self-supervised pretraining and supervised pretraining on ImageNet.
Conclusions: Self-supervised pretraining may boost the performance of deep learning models for dental image analysis when annotated data are sparse.