Evaluation of effectiveness of a virtual AI-based dental assistant in recognizing mixed dentition on panoramic radiographs

Introduction

Artificial intelligence (AI)-based technologies are widely used in different areas, including medicine and dentistry [1]. It is a field of science that aims to enhance data analysis without major human influence. This tool can significantly improve the diagnosis and treatment of many pathologies in different branches of healthcare, such as ophthalmology [2, 3], neurology [4, 5], or cardio­logy [6, 7]. In dentistry, automatic detection can be helpful in orthodontics [8, 9], periodontology [10], restorative and pediatric dentistry [11-13], etc. There are three main types of AI technology: artificial neural networks (ANN), machine learning (ML), and deep learning (DL). Recently, convolutional neural network (CNN) has been popularized, especially in dental and maxillofacial radiology, because it aims at image/pixel recognition. CNN is made of many deeper layers, and it is similar to the human neural network.
One of the platforms based on convolutional neural networks is Diagnocat Inc. (DGNCT LLC, Miami, USA), available online. It is also known as virtual dental assistant. This system allows for uploading intra-oral radiographs (IR), panoramic radiographs (PR), and cone-bean computed tomography (CBCT) scans, and after analysis, it generates a radiographic report. The final result is simple to understand, as the algorithm creates a segmented chart that indicates findings with different colors and sings. The producer claims that Diagnocat can detect over 65 conditions, among which are dental caries, impacted teeth, and periodontal bone loss. Apart from the radiological report, an online platform enables to create CBCT-segmented models.
In order to deliver an accurate diagnosis, it is important to combine dental examination and diagnostic imaging. Panoramic radiograph is a common type of radiological extra-oral image used to visualize anatomical structures, hard and soft tissues pathologies, and even to identify deceased individuals in forensic dentistry [14, 15]. In many branches of dentistry, panoramic radiograph is considered a baseline examination, especially in implantology, surgery, and orthodontics. This radiographic technique has many advantages, including evaluation of both maxillary and mandibular teeth with their adjacent structures, relatively low radiation dose, and it is technically simple and fast to perform [16].

Objectives

The objective of this study was to analyze panoramic radiographs of patients with mixed dentition using Dia­gnocat Inc. software, and to compare the results by a dentist.

Material and methods

Seventy-eight panoramic radiographs of patients with mixed dentition were selected from the database of Department of Dental and Maxillofacial Radiodiagnostics, Medical University, Lublin, Poland, according to inclusion and exclusion criteria. All examinations were performed using a panoramic X-ray unit, VistaPano S with S-Pan technology, with 74 kV, 14.0 mA, and 7 s. Patients data were properly anonymized. Inclusion criteria were as follows: good quality of the image without motion artifacts, patients aged 7-11 years, symmetrical position of the patient, and visibility of both jaws and all teeth. Panoramic radiographs containing ghost shadows, motion artifacts, and with any pathological structures or developmental disorders were excluded. The study group included 35 males (44.87%) and 43 females (55.13%), aged from 7 to 11 years (mean age, 8.6 years). The selected images were assessed for tooth numbering using Viohl system by three fifth-grade students, and then supervised by two dentists with 12 and 5 years of experience in dental and maxillofacial radio­logy. Students’ evaluation time was measured. If dentists agreed with their evaluation, it was considered a ground truth. Subsequently, all panoramic radiographs were analyzed using Diagnocat Inc. software, and the results were compared. Descriptive statistics methods were employed.

Results

The evaluation time for students varied from 2 minutes to 2 minutes and 20 seconds, and Diagnocat analyzed the results in around 2 minutes. The software correctly identified deciduous and permanent teeth in 58 cases (74.35%), while 20 reports (25.64%) of pano­ramic radiographs were analyzed incorrectly. Errors included wrong order of teeth (Figure 1), doubled number of teeth (Figure 2), and failure to identify them (Figure 3). Some of the inaccuracies appeared simultaneously in a single report. In total, there were 48 errors. The overall number of various errors is listed in Table 1.
Th most significant issues occurred in cases involving deciduous teeth with advanced resorption. There were notably more inaccuracies in the assessment of anterior teeth than in the posterior region. However, Diagnocat Inc. software presented 100% effectiveness in identification of third molars, even these in early stages of mineralization. Among all AI-based reports, in the group of females, the percentage of incorrect estimations were 14.1%, and in the group of males, 11.54%. There was a correlation between the number of errors and the age of patients. In 7-year-old children, 80% of panoramic X-rays were assessed incorrectly by Diagnocat Inc. software (Figure 4). In the group of patients over 10.5 years of age, no errors were found.

Discussion

Recently, a large number of studies have focused on the usefulness of artificial intelligence in radiological evaluation [12-14, 16-25], based on different types of radiographic examinations, such as periapical images [17, 18], bitewings [19], panoramic X-rays [14-16, 20-24], and CBCT [12]. Tooth detection and numbering is one of the basic components of appropriate dental exami­nation. AI tools are considered to be helpful and time-saving for charting of radiological images [16, 22]. Tuzoff et al. [24] used 1,574 panoramic radiographs to provide automated tooth detection and numbering based on a convolutional neural network’s method. The results turned out to be very promising. For tooth detection, the CNN architecture achieved a sensitivity of 0.9941 and a precision of 0.9945, while for tooth numbering, the sensitivity was 0.9800. The authors emphasized that the main reasons for errors are overlapped teeth, prosthetic appliances, or too small remaining of dental fragments. Their results are in line with those of the present study, where advanced deciduous tooth resorption significantly decreased the effectiveness of Diagnocat software. To improve recognition of prosthetic appliances by artificial intelligence tools, Ali et al. [25] proposed a novel approach that includes two separate CNN-based detectors, called YOLOv7. One of them detects teeth, and another detects prostheses. The mean average precision (mAP) for prosthesis model was 97.16%, and for teeth detection model, it was 99.06%. The authors claim that prosthetic data added into the process of teeth identification can make the process more accurate. Kılıc et al. [16] used Faster R-CNN Inception v. 2 (COCO) model for automatic detection and numbering of deciduous teeth. The age of patients ranged from 5 to 7 years, which is less than in the current study. The results showed high sensitivity and precision rates, i.e., 0.9804 and 0.9686, respectively. Prados-Privado et al. [21] focused on automatic tooth numbering in a group of adults. Inclusion criteria were panoramic radiographs from patients older than 18 years. The research was based on two main neural network layers: one was employed in object detection and second was used for classification. The method achieved an accuracy level of 93.83%. Kaya et al. [23] performed an evaluation of YOLOv4, a CNN model for both primary and permanent tooth detection and enumeration, based on panoramic radiographs of patients aged from 5 to 13 years. The mean average precision in their study was 92.22%. Zadrożny et al. [22] used Diagnocat Inc. software to assess different anatomical and pathological features. One of them was recognition of missing teeth. The AI-based application resulted in very high sensitivity and specificity rates in this field, i.e., 0.961 and 0.981, respectively. However, the main limitation of this study was a small cohort of evaluated panoramic radiographs (n = 30). Additionally, the authors did not take into account the full report of tooth numbering. Mladenovic et al. [13] presented a case report of six-year-old boy with supernumerary teeth. CBCT scan was performed and transferred into Dia­gnocat Inc. software for analysis. The AI application created a report, identified two supernumerary teeth, and provided a complete treatment plan. The authors claimed that Diagnocat report with segmentation based on 3D examination took approximately 5 minutes to generate.
The limitation of the current research was a relatively small study group and analysis based on panoramic images only. Also, CBCT analysis could be more accurate. Further studies with larger cohorts should be carried out to validate the effectiveness of Diagnocat Inc. software.

Conclusions

The Diagnocat Inc. software demonstrates significant potential in identifying teeth in patients with mixed dentition. AI identification errors are mostly due to the resorption of deciduous teeth. Furthermore, the precision of AI is lower in younger patients. AI tools can assist in faster diagnosis and create potential treatment plans.

Disclosures

1. Institutional review board statement: Not applicable.
2. Assistance with the article: None.
3. Financial support and sponsorship: None.
4. Conflicts of interest: The authors declare no potential conflicts of interest concerning the research, authorship, and/or publication of this article.

References