23 January 2024

---

Artificial intelligence in the diagnosis of gingivitis

Lara Figini

---

Periodontal disease is a chronic inflammatory disease that affects the periodontium and is classified into gingivitis and periodontitis with reversible and irreversible tissue damage.

According to the World Health Organization, the prevalence of periodontal disease is estimated to be more than 50% worldwide and almost a third is represented by severe cases, with clinical attachment loss greater than 5 mm and bone loss greater than 30%. 

Periodontal disease is caused by the accumulation of plaque or biofilm along the gum line, resulting in localized gingival inflammation and ongoing host response. It is difficult for patients to maintain satisfactory plaque control without the continuous supervision of a professional, hygienist or dentist. Artificial intelligence can be used to provide continuous automated visual monitoring and consultation via intra-oral photographs.

There are currently several network architectures used to detect gingivitis via intra-oral photographs with accuracy ranging from 0.47 to 0.83, with 1.00 as the maximum accuracy value. The accuracy of any diagnostic system for clinical use should be as high as possible, and the precision should be at least 0.90 or better.

The aim of the following study was to predict gingival health status with accuracy, in terms of sensitivity and specificity, via a new artificial intelligence system built with DeepLabv3+, after training with an adequate number of intra-oral photographs.

Materials and methods

In a study, published in the International Dental Journal, the authors developed and validated a new artificial intelligence system that can be used to diagnose gingivitis via intra-oral photographs without the intervention of the human eye. The authors collected frontal view intra-oral photographs that met the inclusion criteria. In this study, the artificial intelligence network architecture used was DeepLabv3+, based on Keras (v2.12, Google LLC) with TensorFlow 2 (v2.9, Google LLC). This neural network is highly transferable and offers multiple pre-trained checkpoints to facilitate learning from datasets. Along the gum line, the gingival condition of individual sites was labeled as healthy, diseased, or questionable. Photographs were randomly assigned as training or validation datasets. The training datasets were fed into the new AI system and its accuracy in gingivitis detection including sensitivity, specificity and intersection mean. Accuracy was reported according to the STARD-2015 statement.

Results

A total of 567 intra-oral photographs were collected and recorded, of which 80% were used for training and 20% for validation. As for the training datasets there were a total of 113,745,208 pixels with 9,270,413; 5,711,027; 4,596,612 pixels labeled as healthy, sick, and questionable, respectively. As for the validation datasets, it was 28,319,607 pixels with 1,732,031; 1,866,104; 1,116,493 pixels labeled as healthy, sick, and questionable, respectively.

The AI correctly predicted 1,114,623 healthy and 1,183,718 diseased pixels with a sensitivity of 0.92 and a specificity of 0.94. The average intersection over system union was found to be 0.60 and above the commonly accepted threshold of 0.50.

Conclusions

From the data of this study, which must be confirmed in other similar works, it can be concluded that artificial intelligence could identify specific sites with and without gingival inflammation, with high sensitivity and high specificity on par with human visual examination.