DQ-SAM:Data-Quality–AwareOptimizationforRobustG.V.BlackClassI–IIICaries ClassificationonClinical IntraoralPeriapicalRadiographs

Abstract

Early and reliable classification of dental caries in intraoral periapical radiographs is challenged by heterogeneous image quality in routine clinical workflows. This study develops a data-quality-aware Swin Transformer pipeline that is robust to such variability and evaluates it on a curated single-centre dataset from the SIBAR Institute of Dental Sciences, Guntur. We employ a lightweight Swin-Tiny backbone with a dental-biased transformer block trained using Data-Quality--gated Sharpness-Aware Minimization (DQ-SAM). Each training image is assigned a scalar quality score $q \in [0,1]$, fused from six fast, interpretable cues (Laplacian variance, Tenengrad, tiled RMS contrast, dynamic range, entropy, and edge density) that are robustly normalized within each G.V. Black class. The resulting $q$ modulates both the SAM perturbation radius and the effective AdamW update, emphasizing reliable gradients from higher-quality batches while tempering updates from lower-quality ones. On a three-class periapical test set of 360 radiographs (G.V. Black Classes I--III), the proposed DQ-SAM Swin-Tiny classifier attains $91.9\%$ accuracy with a macro-F1 score of $0.92$ and class-wise ROC--AUC values above $0.97$. Quality-stratified analysis shows stable performance across low-, medium-, and high-quality bands ($\approx 89$--$93\%$ accuracy), indicating robustness to blur and contrast variation. Compared with DenseNet-121 and ResNet-50 baselines trained under the same protocol, the proposed model delivers consistently higher accuracy, particularly in challenging quality regimes. The method supports reliable artificial intelligence for clinical dental imagin