Artificial intelligence-based photographic detection of pink esthetic score attributes using a hybrid deep learning segmentation pipeline: a method development study.

This study aimed to develop and internally validate a novel anatomy-driven artificial intelligence (AI) system for automated postoperative Pink Esthetic Score (PES) evaluation from intraoral photographs. Unlike most existing AI approaches, which rely on end-to-end prediction, the proposed pipeline derives PES attributes from anatomically grounded measurements. A hybrid analytical pipeline integrating instance segmentation and rule-based measurement was developed. Tooth crown segmentation was performed using Mask R-CNN, while gingival segmentation was performed using YOLOv11. Anatomical features were extracted from the segmented structures and converted into ordinal PES attributes through data-driven threshold calibration derived from the training dataset. The reference standard consisted of independent expert scoring of 82 postoperative intraoral photographs. Diagnostic accuracy, confusion matrix analyses, and regression metrics were calculated to evaluate agreement between automated and expert assessments. The hybrid segmentation-based system achieved accuracies of 91.5% for the mesial papilla, 85.4% for the distal papilla, 86.6% for the gingival margin, and 82.9% for gingival color assessment. Exact agreement between automated and expert total PES scores was observed in 57.3% of cases, increasing to 79.3% when a clinically relevant tolerance of ± 1 PES point was applied. Regression analysis demonstrated a mean absolute error of 0.76, a root mean squared error of 1.35, and an R² value of 0.49. An anatomy-driven segmentation framework may provide a transparent and reproducible approach for automated photographic PES attribute assessment in research settings.