Algorithms of artificial intelligence for the evaluation of gingival conditions

Abstract

Introduction. Modern methods of identification and prediction of gingivitis and periodontitis based on artificial intelligence models that analyze intraoral photos or scans data include approaches using algorithms of deep neural networks, method of support vectors, various types of classification trees and logistic regression. However, most studies dedicated to the approbation of such approaches were characterized by a retrospective design and, therefore, associated with limited possibilities regarding translation of the obtained results into the conditions of daily dental clinical practice. Objective of the research. To analyze the available models and algorithms of artificial intelligence that could potentially be used to automatize the process of gingival conditions diagnostics. Materials and methods. The study was organized in the format of targeted literature review with a focused search for data regarding available artificial intelligence models and algorithms developed specifically for the diagnosis and differentiation of gingival diseases. In order to maximize the volume of the primary sample, the search for publications related to the purpose of this study was conducted through the Google Scholar system (https://scholar.google.com/), using the keywords «artificial intelligence», «gingiva» and «periodontology», advanced service functions and filtering of works published in English until April 2024. Results and discussions. The analyzed artificial intelligence models demonstrated the effectiveness of automatic identification for areas of gingivitis and gingival contour disruption with an accuracy level of more than 70% using intraoral digital photographs as the main set of initial data. At the same time, verification accuracy for healthy gingival condition according to the data of the analyzed studies was quite low, which is related with the problems of differentiating segmented areas with reference images of healthy gums, because latter characterized not only by inter-individual, but also by intra-individual variability. The problem of qualitative assessment of gingival conditions using artificial intelligence models based on intraoral scanning data is related to the differences within image acquisition technologies provided by different devices and the so-called effect of «excessive digitization», which leads to distortion within the real representation of soft tissues condition. In addition, the common problem of intraoral scanners in assessing the state of the gingiva is related to the augmented type of graphical surface processing, which is characterized by the absence of pronounced fiduciary markers (physiological or artificial) and the uniformity of planar topological characteristics within the scanned gingival area. Conclusions. Available models and algorithms of artificial intelligence, the purpose of which involves the assessment and differentiation of the gingiva conditions, have demonstrated a high accuracy regarding the automated process of gingivitis cases diagnostics at the patientoriented level (≥ 70%), while the sensitivity of such models for the verification of the healthy gingival condition remains low, and is characterized by a wide range of variation (with the starting point of the range from 0%). The use of artificial intelligence models for the purpose of segmentation and further categorization of gingival areas according to certain clearly defined classes (during the index assessment of the gum condition, or during the initial differentiation and qualitative categorization of the smile type), as well as during the identification of areas with gingival contour disruption and its further possible reconstruction according to the established pattern, is characterized by high values of sensitivity and specificity at the level of the studied samples, however, the clinical validity of these approaches, taking into account the inter- and intra-individual levels of variation for various manifestations of gingival changes, has not yet been proven.