Accuracy of artificial intelligence technology in detecting number of root canals of obturated human maxillary second molars indicated for retreatment : diagnostic accuracy experimental study /

Abstract

Statement of the Problem: Endodontic treatment has always been a challenging procedure and failure to detect all canals and completely disinfecting the root canal system could lead to an endodontic failure. In retreatment cases, the endodontic procedure becomes more challenging and tricky requiring high knowledge and skills. Artificial intelligence software makes it easier in analysis of CBCT scans reducing dentist errors and improving treatment success rates. Aim of the study: The aim of the study is to evaluate the accuracy and reliability of an artificially intelligent software that would analyze, aid, support, and help the endodontist to detect the correct number of root canals of 35 obturated maxillary second molars indicated for retreatment in comparison to both Cone Beam Computed Tomography and Clinical observation. Materials and methods: This study included 35 Patients with Maxillary Second obturated molars indicated for re-treatment who underwent a pre-treatment CBCT, while only a pre-treatment periapical radiograph was taken to aid in access cavity preparation in the clinical stage. The study was levelled into three stages. CBCT Stage: Pre-operative CBCT scans of the patients were taken and randomly assigned to 2 co-supervisors who segmented, interpreted, and recorded the number of canals on a pre-formed information guide. Clinical Stage: A clinical stage where the enrolled patients were randomly distributed upon 6 researchers. Researchers then performed access cavities under DOM. The number of orifices found was recorded in a preformed information guide. AI stage: CBCT scans was uploaded to AI software by researcher and Number of canals found by the software were recorded. The results of the first two stages were then compared to the findings of the third stage to determine software accuracy. Cases with missed canals by the AI software underwent further evaluation of tooth morphological features, to determine the reason for the software’s detection failure. Results: There was a statistically significant difference between detected number of canals by the three methods (P-value = 0.018). AI showed higher percentage of two, three canals and lower percentage of four canals detection. The CBCT stage and the clinical stage showed complete agreement in their findings, however the CNN program inaccurately identified the proper number of canals in three cases. To evaluate the relationship between these cases and the potential impact of morphological variation on the AI software, the canal type was determined using Vertucci's classification. It was found that all three unsuccessful cases had a type II canal. There was an insignificant link between the canal morphology and the AI software's incapacity to detect canals. When the AI program was unable to accurately identify the proper number of canals, the average ratio of canal unity was around 60%. This means that around two-thirds of the canal length was considered as one canal, while only one-third of the length was considered as two separate canals.