Stacked Ensemble Deep Learning Models for Accurate Detection and Size Stratification of Periapical Lesions on Intraoral Radiographs.

Periapical lesions are challenging to detect on intraoral radiographs because of anatomical superimposition and reader variability. This study developed stacked deep learning ensembles for automated detection and radiographic size stratification of periapical lesions. In total, 146 radiographs comprising normal cases and three lesion-size categories were cropped around the root apex and augmented using predefined transformations. Five convolutional neural network backbones were trained, and their probability outputs were combined using multinomial logistic regression (MLR) and gradient boosting (XGBoost) meta-learners. On the internal hold-out test set, both ensembles performed better than the individual backbones. MLR achieved an accuracy of 0.96 and a macro-averaged ROC-AUC of 1.00, while XGBoost achieved an accuracy of 0.95 with a similar AUC. Both ensembles also achieved a sensitivity of 0.94 for lesions < 2 mm. These findings support further investigation of stacking for radiographic decision support, but optimistic estimates from this small, single-centre internal study require prospective multicentre external validation.