Artificial Intelligence for Detecting Periodontitis: Systematic Literature Review

1.1. Background Systematic Review

The development of science and innovation in recent years has resulted in many significant advances in the determination of diagnosis, treatment, and even the determination of preventive measures against periodontal disease [1]. The diagnosis of periodontal disease, which includes damage to soft and hard tissues, should be systematic and should involve clinical examination and radiographic examination [2-5]. Radiographs play an important role in the diagnosis, research, and treatment of periodontal disease, although there are sometimes limitations when periodontal abnormalities are not indicated in radiographic images [6, 7]. The best approach to the diagnosis of periodontal disease is clinical examination and radiographic examination, to establish the diagnosis of periodontal disease or to obtain other supporting evidence for determining the diagnosis [7, 8]. Several types of conventional radiographic examinations can be used to diagnose periodontal disease and even determine prognosis, including periapical, bitewing, and panoramic [2, 6]. Full-mouth periapical radiographs have been the first choice and also the gold standard for evaluating periodontal disease abnormalities [4, 7, 8]. Panoramic radiographs, which capture the entire mouth, have been widely used for their advantages over intra-oral images, such as better patient comfort and easier procedure. The selection of the use of periapical radiograph techniques for the determination of diagnosis should also pay attention to the radiation dose. The recommended recommendation for the radiation dose of one periapical photo is 0.001 mSv, and the recommended radiation dose for panoramic photos is 0.007 mSv [9, 10]. Selection of panoramic radiographs, if necessary, should be combined with periapical radiographs to reduce the amount of radiation that occurs [11]. However, there are often many things that influence consideration in the selection of appropriate radiographic techniques in determining the diagnosis of periodontal disease, especially among clinicians [12, 13].

The use of artificial intelligence in dentistry can be widely applied, depending on needs, ranging from dental emergencies to prosthetic planning. In the use of artificial intelligence, machines will learn from old experiences or data, adapt new data, and process like humans [14, 15]. As the trend of digital systems approaches in dentistry has increased in the last 10-20 years, clinicians lacking accurate diagnostic skills are driving the need for the use of technology, especially Artificial Intelligence (AI) software. This system is believed to reduce costs, time, the need for medical services, and also medical errors that may occur due to human factors [16]. Artificial intelligence intends to reproduce the cognitive processes of the human being and obtain the same result, in this case, the determination of the diagnosis of a disease with that produced by the clinician, with more accuracy and shorter time [17]. In establishing the diagnosis of periodontal disease, AI can be the link between disease indicators, immunology, and microbiology to obtain a diagnosis of periodontal disease [17].

1.2. Research Objectives

The purpose of this systematic review study is to evaluate the use of conventional periapical, bite-wing, and panoramic radiographs that have been an option in helping to establish the diagnosis of periodontal disease, combined with the use of machine learning as artificial intelligence to determine the diagnosis of periodontal disease based on severity parameters periodontal disease according to the 2017 AAP / EFP Workshop, which are pocket depth, clinical attachment loss (CAL) and the pattern and level of alveolar bone damage that occurs. This study reviewed observational research types, such as cross-sectional, cohort, and randomized controlled clinical trials (RCT study), to see the relationship between determining the diagnosis of periodontitis influenced by periodontal disease severity parameters based on periapical and panoramic conventional radiographic examinations. Study indicators are those who are diagnosed with periodontitis and have conventional radiographic supporting examinations accompanied by supporting data that show the severity of the periodontal disease.

2.1. Research Protocol

This systematic review research protocol follows the rules of PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses, which provides research protocol points linked to references used in systematic reviews. This research question uses the PICOS method (Population) for Patients with a diagnosis of periodontitis and aggressive periodontitis aggravated by local predisposing factors, food impaction, food retention, and occlusion trauma (Intervention). The use of machine learning and artificial intelligence in Panoramic photographs in the determination of periodontal diagnosis, (Comparison) Periapical and bitewing radiographic photographs, (Outcomes) CAL, pocket depth, and periodontal bone loss, (Study) Cohort, cross-sectional and randomized control trials.

2.2. Eligibility Criteria

2.3. Literature Search

The literature search for this research includes relevant studies over the past 10 years. The database search engines used in this study were Pubmed and Scopus. The same keyword is used for searches in both databases using the Boolean terms “OR” and “AND.” With the following filters: free full text, humans, English, and MEDLINE.

2.4. Selection of Studi

Two independent researchers (D.F., Y.S.) independently analyzed titles, abstracts, and full text to identify articles eligible for systematic review based on inclusion and exclusion criteria. Disagreements between researchers regarding the selection of articles were resolved by agreement.

2.5. Data Collection

Researchers (D.F) analyze and check data related to research methodology and heterogeneity of clinical parameters. Analysis of data from each article as follows: research design studies (cohort, cross-sectional, or randomized clinical trial), characteristics of participant studies, such as diagnosed periodontitis, local predisposing factors that affect periodontitis, the use of artificial intelligence in reading conventional radiographic image data and quantitative data (parameters used (outcome)), such as Clinical Attachment Loss (CAL), alveolar/periodontal bone loss and pocket depth. Use sensitivity and specificity values for each existing quantitative data (outcome).

3.1. Determination of Research Design Studies and Research Characteristics

The total number of research studies that fit the inclusion criteria of this system literature review research obtained from the journal article search database (Pubmed and Scopus) is 32 journal articles published between 2011 and 2021 (December 18, 2011, to December 17, 2021). After passing the examination which includes the title, abstract, free full text, and language used, then only 2 journal articles meet the criteria, as shown in the PRISMA chart below [18, 19] (Fig. 1).

The detailed characteristics of the 2 research studies that are eligible for the criteria appear in the following table, accompanied by researchers in each research study, year of publication, time of study, research design, and parameters used as quantitative data.

3.2. Results from Each Research Study

Research conducted by Kim et al. [20] and Krois et al. [21], (Tables 2 and 3) aimed to determine the severity of periodontal disease with periodontal bone damage parameters using panoramic radiographs analyzed by machine learning, then compared with the determination of periodontal bone loss, previously marked by clinicians who are observers, whose reliability will be assessed intra- and inter-rater between observers (Table 4) [20, 21]. To the best of our knowledge, determining periodontal bone damage on dental radiographs is often difficult due to the limitations of different clinicians for damage determination periodontal bone [20-23]. Kim et al. [20] used a Deep Neural Transferred Network (DeNTNet), which is a development of the Convolutional Neural Network (CNN) used in the research conducted by Krois et al. [21]

For the determination of bone damage (Table 4), Krois et al. [21] used 3 parameters, there are cemento enamel junction (CEJ) in the mesial and distal regions of the tooth, the most apical point of the apex of the tooth, and the most apical point of extension loss of the apex of the alveolar bone, then determined the percentage value of bone damage, which calculates the distance from CEJ to the top of the alveolar bone divided by the distance from CEJ to the dental apex [21], if the tooth is covered by the restoration, the most apical point of the restoration is taken as CEJ [24]. For premolar and molar, which have two or more dental roots, the largest percentage is taken in the distal or mesial region. The use of periodontal bone loss percentages instead of precise calculations, such as millimeters, helps reduce radiographic film distortion caused by the position of the patient and magnification of the panoramic plane [21]. While Kim et al determine bone damage, based on the presence or absence of alveolar bone damage [20]. Both studies used clinicians with sufficient practical experience as dentists, periodontists, and radiologists [20, 21].

Table 4 shows the agreement among the clinicians. There were six dentists in the study conducted by Krois et al. [21]; two of them specialized in operative dentistry and periodontology, and four of them were general dentists [21]. The score of Kappa (0,52 as a moderate’s score) determined the intra-rater reliability of this study [25, 26]. In the study conducted by Kim et al. [20], the determination of bone damage began with a pre-trained model, namely the determination of ROI (region of interest) in panoramic, then the determination of ROI in the trained model to predict periodontal bone loss (PBL) on each tooth. Then, the premolars and molars are made a separate assessment on the existing trained model, and this is due to the morphology of the two types of teeth, which sometimes cannot be seen in panoramic photos, which could need the training data set twice, compared to anterior teeth (incisive and canine) which have one root, this is because, in research conducted by Kim et al., ROI segmented for premolar and molar are performed vertical split [20]. The image data will be adjusted up to 512x512 pixels. While in research conducted by Krois et al [21], image data will be processed digitally, previously it will be converted to gray-scale, and then teeth on the upper jaw will be flipped in the next process. Then, the image data is adjusted to 64x64 pixels to facilitate data reading. Randomly, the data set will be divided into two data sets: a training data set and a validation data set [21]. The results of research conducted by Kim et al. and Krois et al. with diagnostic tests produced sensitivity values with a range of 0.70 to 0.90 and specificity of 0.46 up to 0.96 (Table 5) [20, 21].

The assessment of diagnostic test results based on bone damage in each segment of the dentition can be seen in Table 6 below.

Accuracy in the anterior region (incisive and canine) is lower than in the posterior region (premolars and molars). Radiographic interpretation of the anterior region is often influenced by the speed and inclination of the rotation of the panoramic plane, affecting the radiographic image of the anterior region and giving the picture that the anterior region is narrower than the posterior region [27].