Segmentation of Dental Restorations on Panoramic Radiographs Using Deep Learning

Objectives: Deep Convolutional Neural Networks such as U-Net have been widely used for medical image segmentation. Dental restorations are prominent features of dental radiographs. Applying U-Net on the full panoramic image is challenging, as the shape, size and frequency of different restoration types vary significantly. It was hypothesized that models trained on smaller equally spaced rectangular image crops (tiles) of the full panoramic would outperform models trained on the full panoramic scan.
Methods: The dataset consisted of 1781 panoramic radiographs. Fillings, crowns, root canal treatments, and implants were segmented pixelwise by dental experts. The radiographs were cropped into different number of tiles for training. We used U-Net architecture pretrained on ImageNet. The data set was randomly split into a training (70%), validation (20%), and test set (10%), for fine-tuning, model selection and evaluation, respectively. The Dice loss function with adaptive learning rate, early stopping and data augmentation was used for model training.
Results: Training with an increased number of tiles improved the model performance and reduced the time of model convergence. The F1-score for the full panoramic image was 0.68, compared to 0.80, 0.90 and 0.94 for 6, 10 and 20 tiles, respectively. For root canals treatments, which are small, cone-shaped features that appear less frequent on the radiographs, the performance improvement was considerable.
Conclusions: Semantic segmentation models trained on panoramic radiographs are biased towards the more frequent and extended classes, hence, reducing the accuracy of the classifier. Training on tiles of panoramic radiographs and pooling the individual results thereafter, improved classification performance and reduced time to model convergence for segmenting dental restorations.