Robust classification of dental tissues using LIBS data based on data reconstruction and distance prior strategy

Modern dental procedures primarily rely on visual inspection and tactile feedback to distinguish enamel from dentin, which may result in irreversible loss of healthy tissue due to excessive preparation. Laser-induced breakdown spectroscopy (LIBS) offers potential for intraoperative tissue identification, but its single-pulse signals are affected by laser energy fluctuations and dental-tissue matrix effects. This study proposes a robust LIBS-based classification method integrating interactive spectral reconstruction with a hydroxyapatite (HAP) spectral-distance prior. A multi-scale sliding-window reconstruction strategy was developed to organize same-site spectral pair data acquired from interval breakdowns into multi-channel matrices, thereby capturing both their shared spectral characteristics and fluctuation differences. Using the mean spectrum of standard HAP as a static reference, Euclidean-distance and Voigt-fitting-based distance features were extracted to quantify physicochemical deviations from HAP and reduce interference caused by matrix effects. A Dual-Dimension Attention Spectral Fusion Network (DDA-Net) was then constructed for enamel-dentin classification. DDA-Net outperformed other baseline models under mixed-data splitting (Mode 1), cross-time splitting (Mode 2), and cross-sample validation. With the Euclidean-distance strategy, it achieved an overall accuracy of 94.88% in Mode 1. In the more challenging Mode 2, its accuracy exceeded those of conventional baselines by 7 . 7 5 % − 1 7 . 8 5 % . In leave-one-tooth-out cross-validation, DDA-Net achieved 90.33% accuracy on unseen teeth and exceeded the three backbone-replacement variants by 1 . 9 1 % − 7 . 3 3 % under identical input conditions. SHAP analysis indicated that the model formed a nonlinear decision mechanism based on coordinated multi-band spectral information. These results demonstrate the potential of the proposed method for accurate and robust intraoperative discrimination between enamel and dentin.