Clinician-Led Development and Feasibility of a Neural Network for Assessing 3D Dental Cavity Preparations Assisted by Conversational AI.

Abstract

<b>Introduction:</b> Artificial intelligence is emerging in dental education, but its use in preclinical assessment remains limited. Large language models like ChatGPT^® V4.5 enable non-programmers to build AI models through real-time guidance, addressing the coding barrier. <b>Aim:</b> This study aims to empower clinician-led, low-cost, AI-driven assessment models in preclinical restorative dentistry and to evaluate the technical feasibility of using a neural network to score 3D cavity preparations. <b>Methods:</b> Twenty mandibular molars (tooth 46), each with two carious lesions, were prepared and scored by two expert examiners using a 20-point rubric. The teeth were scanned with a Medit i700^® and exported as .OBJ files. Using Open3D, the models were processed into point clouds. With ChatGPT's guidance, the clinician built a PointNet-based neural model in PyTorch, training it on 20 cases and testing it on 10 unseen preparations. <b>Results:</b> In training, the model achieved an MAE of 0.82, RMSE of 1.02, and Pearson's r = 0.88, with 66.7% and 93.3% of the predictions within ±5% and ±10% of the examiner scores, respectively. On the test set, it achieved an MAE of 0.97, RMSE of 1.16, and r = 0.92, with 50% and 100% of scores within ±5% and ±10%, respectively. These results show a strong alignment with examiner scores and an early generalizability for scoring preclinical cavity preparations. <b>Conclusions:</b> This study confirms the feasibility of clinician-led, low-cost AI development for 3D cavity assessment using ChatGPT, even without prior coding expertise.