Accuracy of artificial intelligence technology in detecting the number of root canals in human mandibular first molars : A diagnostic accuracy experimental study /

Abstract

Artificial intelligence (AI) is revolutionizing various fields of medicine, including dentistry. In endodontics, AI has the potential to improve diagnostic accuracy, treatment planning, and overall patient care. There is minimal scientific evidence concerning the accuracy of the detection of the number of canals using Artificial Intelligence. The aim of this study is to define the accuracy of an Artificial Intelligence (AI) Software to detect the number of root canals in the mandibular first molars compared to the standard clinical method under magnification and to the gold standard using a Cone Beam Computed Tomography (CBCT). In addition, this study aims to determine the morphological variations in cases where the AI failed to detect all the canals. Thirty-Five patients presented to MIU dental clinics with mandibular first permanent molars requiring root canal treatment were selected according to the eligibility criteria. Teeth that met the criteria had pre-operative periapical x-rays taken. Adequate case and medical history were taken. Detection of the number of root canals was achieved by 3 methods: Radiographically using the CBCT as the gold standard, clinically using the DOM and using an AI software. After signing the informed consent, the patients underwent CBCT imaging using Endo mode to minimize the radiation exposure before treatment initiation. CBCT was coded based on the patients file numbers instead of names. The number of canals identified was recorded in a pre formed information guide. Using randomization software, the 35 patients were divided up amongst 6 post graduate students at random. A dental operating microscope was used to prepare the access cavity for each patient. In order to minimize the radiation dose to the patient, the CBCT was given to the postgraduate student after the number of canals was recorded. This helped to prevent inter-treatment periapical radiographs as working length X-rays. After that, the principal researcher, who was blind to the outcomes of the CBCT stage, carried out the AI stage individually. The Diagnocat AI Software was used to upload the CBCT images, and CBCT segmentation and deep learning techniques were employed. Next, the software's total number of canal identifications was recorded. The gathered data was divided into three groups based on the canal detection method into: CBCT Findings, Clinical Findings and AI Findings. Cases containing canals that were missed by the AI software had additional examination of the morphological features of each tooth from the CBCT. The six cases in which the AI software successfully detected the number of canals were randomly selected as a comparative group. For both groups the following features were included: the inter-orifice distance was measured in order to accomplish this, the length of the root, Vertucci classification type and the length of the Canal division measured from the apex and CEJ. In addition, from the axial cut of the CBCT, the inter-orifice distance was measured in the same manner between the canals that were not detected by the AI. Afterward, the mean inter-orifice distance was calculated for both the failed and successful cases. Those measurements were utilized to calculate the percentage of canal unity within the entire root length associated to the missed canals. The degree of accuracy was 100% for the DOM, 100% for the CBCT and 82.86% for the AI. It was clear that application of AI software didn’t detect the exact number of canals in all cases. The AI software detected the correct number of canals in only 29 cases. The use of DOM and CBCT detected all canals in all the 35 cases. Another analysis was done to relate the inaccuracies of the software to the morphology of the wrong cases detected. It was evident that the inter-orifice distance for the incorrectly detected cases was less than that of the cases correctly detected. Added to that, the percentage of canal unity was analyzed for the distal canals, which is the measure of the unity of both distal canals along the entire length of the root. The greater the union between both distal canals, the less likely the AI software will distinguish them as separate canals. The average percentage of canal unity of the incorrectly detected cases was greater than 50% and the percentage of canal unity for the correctly detected cases was a maximum of 20% which allowed the software to detect two distal canals.