Ensemble learning for predicting periapical lesions from dental radiographs

Diagnosing periapical cysts and granulomas using periapical radiographs is challenging due to subtle radiographic differences. This study aimed to develop a machine learning-based diagnostic framework to improve lesion classification through ensemble learning and dentin standardization.

Five models-Support Vector Classifier (SVC), Nu-SVC, K-Nearest Neighbors (KNN) and Decision Tree-were trained on 144 pre-treatment periapical radiographs (70 cysts and 74 granulomas). Dentin standardization normalized grayscale values to ensure feature consistency. A weighted soft voting strategy was employed to integrate predictions, with model weights derived from individual diagnostic performance. Model performance was evaluated via five-fold cross-validation. Statistical significance among models was assessed using the Friedman test, followed by Wilcoxon signed-rank tests for pairwise comparisons (α = 0.05).

The ensemble method achieved a precision of 0.83 and a sensitivity of 0.83, demonstrating robust diagnostic performance. Statistical analysis revealed significant differences among models in specificity (P = 0.001) and negative predictive value (P = 0.044), with the ensemble method reaching a peak specificity of 0.83. Although numerous improvements were observed in precision and sensitivity compared to several base models, these differences did not reach statistical significance (P > 0.05). Overall, the ensemble method demonstrated balanced performance across metrics, although its improvements over individual models were not statistically significant.

By integrating dentin standardization with a weighted ensemble approach, the proposed method provides a reliable, non-invasive tool for improving radiographic differentiation of periapical cysts and granulomas. These results support the potential of intelligent diagnostic systems in dental radiology.