IADR Abstract Archives
Application of Artificial Intelligence for Automated Tooth Segmentation and Cavitated Caries Detection on Colored Images
Objectives: This study aims to demonstrate that a deep learning system can identify cavitated caries on individual anterior teeth from digital RGB images.
Methods: Digital RGB images of anterior upper and lower teeth were acquired from 264 subjects aged 11 - 12 using a Nikon D5600 DSLR with a Medical Nikkor macro lens. A polygon around each visible tooth and ICDAS Score 4 – 6 primary dental caries lesions were drawn by dental students using the LabelMe image annotation software. Each anterior-facing tooth was labeled with its individual tooth number, giving 12 unique tooth classes for prediction. Cavitated caries were all labeled with a 13th caries class.
A Mask-RCNN instance segmentation model was trained on a randomly selected subset of 210 annotated images using the python packages PixelLib, and Keras based on TensorFlow 2.0. Hyperparameter tuning was performed using a validation set of 54 images with the help of the Weights and Biases online tool. The final model was used to automatically detect teeth and caries on a hold-out test set of 93 images, and these automatically-detected polygons were measured against the hand-drawn polygons.
Results: The trained model was shown to have an Intersection over Union (IOU) of 0.350 for 12 class teeth segmentation and 0.656 for caries detection.
Conclusions: An Artificial Intelligence program was developed to identify anterior teeth and detect cavitated caries of colored images. There remains room for improvement, but our research demonstrates that adequate detection performance can be achieved with a relatively small set of annotated images. This tool could be used for epidemiological purposes, tele-dentistry, and forensics.
Methods: Digital RGB images of anterior upper and lower teeth were acquired from 264 subjects aged 11 - 12 using a Nikon D5600 DSLR with a Medical Nikkor macro lens. A polygon around each visible tooth and ICDAS Score 4 – 6 primary dental caries lesions were drawn by dental students using the LabelMe image annotation software. Each anterior-facing tooth was labeled with its individual tooth number, giving 12 unique tooth classes for prediction. Cavitated caries were all labeled with a 13th caries class.
A Mask-RCNN instance segmentation model was trained on a randomly selected subset of 210 annotated images using the python packages PixelLib, and Keras based on TensorFlow 2.0. Hyperparameter tuning was performed using a validation set of 54 images with the help of the Weights and Biases online tool. The final model was used to automatically detect teeth and caries on a hold-out test set of 93 images, and these automatically-detected polygons were measured against the hand-drawn polygons.
Results: The trained model was shown to have an Intersection over Union (IOU) of 0.350 for 12 class teeth segmentation and 0.656 for caries detection.
Conclusions: An Artificial Intelligence program was developed to identify anterior teeth and detect cavitated caries of colored images. There remains room for improvement, but our research demonstrates that adequate detection performance can be achieved with a relatively small set of annotated images. This tool could be used for epidemiological purposes, tele-dentistry, and forensics.