341 - From Image to Chairside Decision-Making: Unsupervised Morphological Clustering of Carious Lesions in Intraoral Photographs and Its Alignment with with Clinical Severity

Background: Most deep-learning approaches for detecting dental caries from intraoral photographs focus on binary classification or lesion localization. However, little attention has been given to the morphological diversity of carious lesions or how image-derived phenotypes relate to established clinical severity scales such as ICDAS. This study investigates whether unsupervised learning can identify clinically meaningful caries morphologies using publicly available intraoral images.

Methods: We analyzed the Annotated Intraoral Image Dataset for Dental Caries Detection, comprising 6,313 images with annotated lesion bounding boxes. From 2,150 caries-positive images, 6,687 lesion-level patches were extracted. Twenty-four handcrafted features were computed for each patch, including color (LAB statistics), morphology (bounding-box dimensions and area), and texture (local binary pattern descriptors). After feature standardization and dimensionality reduction using principal component analysis, k-means clustering was performed, with cluster number selected using multiple validation indices. Feature contributions were assessed using a post-hoc random forest model. Two calibrated dentists independently reviewed lesion montages, assigned clinical severity (D1/D2), estimated ICDAS equivalence, and recommended management strategies.

Results: Two stable morphological clusters were identified. A smaller cluster (10.5%) consisted of larger, darker, and more irregular lesions, corresponding visually to early surface-breaking lesions and associated with ICDAS ≈3 and preventive or monitoring approaches. The larger cluster (89.5%) showed smoother, lighter lesions consistent with cavitated dentin lesions (ICDAS ≈5) typically requiring restorative treatment. Lesion size accounted for most cluster separability, with color and texture providing modest additional discrimination. Alignment between clusters and clinical categories was partial, reflecting a continuous morphological spectrum.

Conclusion: Unsupervised clustering of intraoral photographs reveals physiologically plausible caries phenotypes that broadly align with clinical severity and treatment decisions, supporting morphology-aware image-based decision support in dentistry.