Skip to main content

Artificial intelligence in the detection and classification of dental caries


“Dental caries is one of the most common infectious chronic diseases, affecting half of the world’s population. The accurate and early diagnosis of dental caries is fundamental to determining the most appropriate treatment. Dental caries is commonly diagnosed clinically by visual-tactile assessment and radiographic examinations. Additional devices that have been used to detect dental caries include digital imaging, light-emitting diode technology, fiber-optic transillumination, and fluorescence cameras and lasers”

In a study conducted by Ahmed et al., published in The Journal of Prosthetic Dentistry (Online Version), and titled ‘Artificial intelligence in the detection and classification of dental caries’, researchers utilized advanced technology to improve the detection of dental caries, or cavities, in patients. They collected and analyzed bitewing radiographs. These radiographs were processed using a software program that segmented and anonymized the images for their analysis. The researchers employed supervised learning algorithms trained on segmentation tasks to identify and classify carious lesions based on the modified King Abdulaziz University classification. They utilized popular deep learning models – ResNet50, ResNext101, and Vgg19 – as encoders, all pretrained on ImageNet weights. Through the combination of multiple models via ensemble learning, they aimed to create a more robust and accurate caries detection model.

The evaluation of the model’s performance was based on two main metrics: the mean score for intersection over union (IoU) and the F1 score. The IoU measures the similarity between the predicted and ground truth areas, while the F1 score combines precision and recall into a single metric. These measurements were used instead of accuracy due to imbalances observed in datasets.

Results from the study showed promising outcomes. The model achieved a mean IoU score of 0.55 proximal carious lesions on a 5-category segmentation assignment and an F1 score of 0.535 using 554 training samples. Additionally, the segmentation model displayed a sensitivity of 0.76, precision of 0.87, and an F1 score of 0.81.

Comparison tests revealed that the AI models outperformed an assistant professor of dentomaxillofacial radiology with 2 years of experience and an assistant professor of restorative dentistry with 3 years of experience when assessing F1 scores.

However, the study acknowledged certain limitations. In cases where radiographs were overexposed or underexposed, the model tended to mislabel these areas as artifacts. Additionally, overlapping proximal surfaces presented challenges for accurate labeling. The size and information limitations of the dataset’s lowest truth labels also posed constraints on the model’s performance. To enhance early detection of dental caries, it would be necessary to collect a larger and more diverse set of training samples. Overall, this study validates the potential of developing an accurate car detection model that can expedite caries identification, improve clinician decision-making, and enhance the quality of patient care. With further research and improvements, this technology holds great promise for enhancing dental diagnostics and promoting early intervention in oral health.


Related links:

Mahmoud H. Al-Johani

Author Mahmoud H. Al-Johani

More posts by Mahmoud H. Al-Johani

Leave a Reply