Prediction of Caries Lesion Surfaces on Bitewings Using Deep Learning

Objectives: Deep Learning using Convolutional Neural Networks (CNNs) has been used for caries detection on bitewing radiographs of permanent teeth. So far, CNNs do not provide information on the lesion surfaces (distal, mesial or occlusal), which would be beneficial for reporting of findings. We applied a CNN aimed to assess the discriminatory ability for caries lesion localization on bitewing radiographs of permanent teeth.
Methods: The dataset contained masks for each tooth, restorations and caries lesions generated in a previous study from 3775 bitewing radiographs of permanent teeth. The corresponding caries lesion surfaces were labeled by dental experts for each image. Mesial and distal labels were translated to left and right to abstract complexity. Data was split randomly into training (70%), validation (20%) and test (10%) sets, stratified by labels (distribution: left: 41.6%; right: 40.1%; occlusal 18.3%). A ResNet-34 classification model pre-trained on ImageNet with the AdamW optimizer (lr=0.0001) was trained up to 50 epochs with a batch size of 32. To prevent overfitting, dropout and data augmentation methods such as horizontal flipping and image rotation were applied.
Results: The model achieved a high accuracy (total: 97.9%; left: 98.7%; right: 99.3% occlusal: 92.8%) and F1-score (total: 0.98; left: 0.98; right: 0.99; occlusal: 0.94) on the test set.
Conclusions: A ResNet-34 classification model trained on a limited amount of data showed high discriminatory ability to classify caries lesion surface locations on bitewing radiographs of permanent teeth.