Discovering knowledge in panoramic radiographs with XAI: a Thai youth study (aged 7–25)

Abstract We present Knowledge XAIeve, a framework for discovering knowledge from convolutional neural network (CNN) models trained on panoramic radiographs, translating a model’s internal reasoning into explicit, testable hypotheses. The framework operates through five stages: (1) preparing the dataset, (2) training a task-specific CNN model, (3) applying global explainability to localize influential semantic regions, (4) performing systematic perturbation to quantify their influence, and (5) statistically validating the resulting patterns using linear mixed-effects models. We demonstrated its application using 5132 panoramic radiographs from 2778 Thai patients aged 7–25 years across two scenarios: sex classification (accuracy 87.38%, sensitivity 87.61%, specificity 87.16%) and age estimation (RMSE 1.96 years). The framework reaffirmed the lower third molar as the primary predictor for age estimation, revealing a nonlinear, sex-dependent relationship between the crown–root ratio and predicted age ( p < 0.001), and reaffirmed the upper canine while discovering upper third molar as key predictors for male and female sex classification, respectively ( p < 0.001). Follow-up validation confirmed that sexual dimorphism in the upper third molar is driven primarily by root rather than crown dimensions. These findings demonstrate the potential of Knowledge XAIeve for generating novel, biologically grounded hypotheses from CNN models trained on panoramic radiographs.