Dental caries detection using a semi-supervised learning approach

Early diagnosis of dental caries progression can prevent invasive treatment and enable preventive treatment. In this regard, dental radiography is a widely used tool to capture dental visuals that are used for the detection and diagnosis of caries. Different deep learning (DL) techniques have been used to automatically analyse dental images for caries detection. However, most of these techniques require large-scale annotated data to train DL models. On the other hand, in clinical settings, such medical images are scarcely available and annotations are costly and time-consuming. To this end, we present an efficient self-training-based method for caries detection and segmentation that leverages a small set of labelled images for training the teacher model and a large collection of unlabelled images for training the student model. We also propose to use centroid cropped images of the caries region and different augmentation techniques for the training of self-supervised models that provide computational and performance gains as compared to fully supervised learning and standard self-supervised learning methods. We present a fully labelled dental radiographic dataset of 141 images that are used for the evaluation of baseline and proposed models. Our proposed self-supervised learning strategy has provided performance improvement of approximately 6% and 3% in terms of average pixel accuracy and mean intersection over union, respectively as compared to standard self-supervised learning. Data and code will be made available to facilitate future research.

Automated summary

This article proposes an efficient self-training method for caries detection and segmentation, using a small set of labelled images to train the teacher model and a large collection of unlabelled images to train the student model. By using centroid cropped images of the caries region and different augmentation techniques for self-supervised training, computational and performance gains are achieved compared to fully supervised learning and standard self-supervised learning methods. Evaluation results show that our proposed self-supervised learning strategy improves average pixel accuracy and mean intersection over union by approximately 6% and 3%, respectively, compared to standard self-supervised learning.