Deep Learning Enhances Failure Mode Analysis in Dental Adhesion Studies

Abstract

Failure mode analysis after shear bond strength testing is essential for evaluating adhesive performance yet remains highly subjective when relying solely on optical microscopy. This study aimed to develop and evaluate a convolutional neural network (CNN) for automated failure mode classification after shear bond strength using optical microscopy images, trained on ground truth derived from focus variation profilometry supported by scanning electron microscopy. A dataset of 434 fractured interfaces from shear bond strength tests on human teeth was labeled with optical profilometry validated by scanning electron microscopy. Optical microscopy images were acquired at 100× magnification. A CNN was trained exclusively on optical images to classify failure modes and compared with expert human observers. Classification performance was assessed per the weighted F1 score. The CNN achieved a mean ± SD of 0.893 ± 0.032 across 20 folds. Individual expert observers obtained F1 scores of 0.895, 0.911, and 0.824. When aggregated into a consensus reference, expert agreement reached an F1 score of 0.914. Misclassifications predominantly involved specimens with failure mode proportions close to the predefined threshold separating mixed from predominantly adhesive failures. A CNN trained on profilometry-based ground truth achieved expert-level performance in classifying failure modes from optical microscopy images. This automated approach reduces operator-dependent variability and offers a scalable objective alternative to visual classification, representing a significant advance toward standardizing failure mode analysis in dental adhesion research.