Accuracy of artificial intelligence technology in detecting root canal number and morphology in maxillary second molars : A diagnostic accuracy experimental study /

Abstract

A comprehensive understanding of root canal space anatomy is essential for clinicians to be able to treat and prevent apical periodontitis, the main reason for endodontic therapy. Approximately 40% of failed endodontic therapy is due to missed canals, indicating lack of clinical thoroughness in treatment. Maxillary second molars present with varying internal morphologies in different individuals, which is affected by age, genetics, and gender. Inadequacy in detecting these variations during treatment is reflected by the upper molars showing the top rates of endodontic treatment failures. Although many techniques have previously been employed to facilitate canal detection, CBCT is currently considered the gold standard in clinical canal structure exploration, as it has the potential to change initial diagnoses and treatment plans in almost 50% of cases. However, due to problems such as interobserver differences in interpreting scans and the need for training clinicians to read CBCT scans, a special interest has emerged in artificial intelligence (AI) technology to help overcome these problems. In this study, the accuracy of AI software in detection of canal number was compared to the radiographic detection using CBCT and the clinical detection using a DOM in maxillary second molars. Thirty-five patients with maxillary second molars requiring primary RCT with signs and symptoms of irreversible pulpitis or necrotic teeth that are restorable were selected. The number of canals were first detected radiographically from the CBCT by experienced endodontists and recorded, then 6 postgraduate students performed an EAC under the DOM and ultrasonic troughing and recorded canal number detected. Finally, the CBCT was uploaded to the AI software to detect the canal number. AI software accuracy was determined by comparing its results to the results of the 2 other groups. Further analysis of the AI generated report was done to assess the cases that canal number was incorrectly detected to evaluate the possible morphological features that might have been the reason for the incorrect detection. Another six cases in which AI software successfully detected the number of canals were randomly selected as a comparative group. Four main features were additionally assessed thoroughly, namely, the inter orifice distance, length of the mesiobuccal root, level of canal division from the root apex, and canal anatomy based on Vertucci classification. The method of analysis of the CBCT for data extraction for this assessment was standardized for all cases and done at the level of the CEJ. Results showed that the CBCT showed the highest percentage accuracy (100%) in the interpretation of canal morphology and number, while the clinical detection followed closely (94.2%), and the AI software was the lowest percentage accuracy (82.8%) amongst the 3 modalities. Its level of confidence was generally high with 79% of the correct cases and 50% of the incorrect cases being in the highest confidence level respectively. The four additional features assessed showed that the average inter-orifice distance is lower in the incorrect group and the average percentage canal unity was much higher. This indicated that the software is less likely to detect the MB2 canal if it is unified with the MB1 for more than 50% of the entire canal length, and that the smaller the inter-orifice distance the less likely is the software in detecting the MB2 canal. The canals’ Vertucci type did not seem to influence the software’s ability to detect the canal number. In general, the results of the AI software were promising, even if it is currently less than the other 2 modalities used, since the nature of the AI software with deep learning strategies spontaneously improves with the greater amount of data input to it with time. Although many studies agreed with the results of this study regarding the percentage accuracy of AI software, however some studies showed differing results that demonstrated higher accuracy of the AI software. This could be attributed to the differences in the study design, the type of AI software used, and the sample size. However, all studies agree that the future of AI software use in endodontic canal detection is highly encouraging.