Deep Learning-Driven Intraoperative Assessment of Pulp Stumps for Precision Pulpotomy.

OBJECTIVES: To develop and validate an automated artificial intelligence (AI) assessment framework based on the self-supervised DINOv2 vision transformer architecture, which objectively classifies intraoperative pulp stump images as "suitable" or "unsuitable" for capping during pulpotomy. METHODS: An AI model utilizing the DINOv2 architecture was trained on 443 high-resolution microscopic images of pulp stumps. The model's diagnostic performance was evaluated using an independent internal test set (n=93) and an external test dataset consisting of images from the literature (n=21). A human-machine comparative analysis was conducted across both test sets to evaluate the AI model's accuracy, sensitivity, and specificity against those of expert endodontists and novice general practitioners. RESULTS: The DINOv2 model achieved a cross-validation accuracy of 0.9398, significantly outperforming the classical supervised convolutional neural networks. The finalized model demonstrated high accuracy on the internal test set (0.9570) and encouraging performance on the external test set (0.8571). In the comparative analyses, the AI model's diagnostic accuracy was comparable to that of expert endodontists (p > 0.05) and significantly superior to that of novices (p < 0.05). Novices tended to over-diagnose capping suitability, demonstrating high sensitivity but notably low specificity (0.3750-0.5000). Conversely, the AI model maintained a balanced performance with high specificity. CONCLUSIONS: The self-supervised DINOv2 framework provides a promising proof-of-concept tool that accurately approximates expert visual judgement for intraoperative pulp stump assessment. By bridging the clinical experience gap and preventing overdiagnosis, this decision-support system holds substantial promise for enhancing the precision and predictability of endodontic treatments.