IADR Abstract Archives
Artificial Intelligence for Tooth Segmentation and Numbering on Panoramic Radiographs
Objectives: The main goal of this study is to use machine learning tools to aid dental specialists in tooth analysis by automating the process of teeth segmentation and numbering. That is detecting each tooth, determining its type, and knowing its shape and boundaries using deep convolutional neural networks (CNN).
Methods: We used a total of 543 panoramic radiographs. We dedicated 435 panoramic radiographs for training. We used 5-fold cross-validation and tested on 108 panoramic radiographs. Method: This study addressed the problem of automatic tooth segmentation on panoramic radiographs following an instance segmentation approach. We employed Mask-Region-based CNN (Mask-RCNN) to detect, segment, and classify each tooth in a panoramic radiograph.
Results: Our model achieved competitive results in tooth segmentation and numbering compared to recent studies. In the tooth numbering task, the proposed model achieved an Average Precision calculated at Intersection-over-Union (IoU) of 0.5 (AP50) of 98.6, and AP75 of 90. In tooth segmentation, i.e tooth shape, our model achieved AP50 of 98.3 and AP75 of 88.3
Conclusions: We addressed the problem of tooth segmentation and numbering by employing an end-to-end deep learning CNN. Our results showed that CNN can be used for automatic tooth segmentation and numbering on panoramic radiographs. The proposed method showed an improved average precision score compared to recent studies. Future work should focus on improving the model robustness and generalization by adding more panoramic radiographs from different machines and including images of teeth with fractures, abnormalities, or infections.
Methods: We used a total of 543 panoramic radiographs. We dedicated 435 panoramic radiographs for training. We used 5-fold cross-validation and tested on 108 panoramic radiographs. Method: This study addressed the problem of automatic tooth segmentation on panoramic radiographs following an instance segmentation approach. We employed Mask-Region-based CNN (Mask-RCNN) to detect, segment, and classify each tooth in a panoramic radiograph.
Results: Our model achieved competitive results in tooth segmentation and numbering compared to recent studies. In the tooth numbering task, the proposed model achieved an Average Precision calculated at Intersection-over-Union (IoU) of 0.5 (AP50) of 98.6, and AP75 of 90. In tooth segmentation, i.e tooth shape, our model achieved AP50 of 98.3 and AP75 of 88.3
Conclusions: We addressed the problem of tooth segmentation and numbering by employing an end-to-end deep learning CNN. Our results showed that CNN can be used for automatic tooth segmentation and numbering on panoramic radiographs. The proposed method showed an improved average precision score compared to recent studies. Future work should focus on improving the model robustness and generalization by adding more panoramic radiographs from different machines and including images of teeth with fractures, abnormalities, or infections.