Multi-modal lesion classification in dental radiographs using Cnn–radiomics fusion with explainable ensemble learning

Periapical radiography is a fundamental imaging method for detecting root lesions; however, interpretation variability and the possibility of missing small pathologies limit diagnostic accuracy. In this study, we developed an explainable, multimodal AI framework that integrates deep features extracted from convolutional neural networks (CNNs) with radiomic features computed solely from lesion regions identified by Grad-CAM maps. A total of 2,285 periapical radiographs were processed using six CNN architectures, including EfficientNet-B1/B4/V2M/V2S, ResNet-50, and Xception. The deep embeddings obtained from these networks were combined with first-order, GLCM, GLRLM, GLDM, NGTDM, and shape-based radiomic descriptors calculated within the Grad-CAM mask. The combined feature vectors were reduced in dimensionality using PCA or SelectKBest and then classified with Random Forest and XGBoost algorithms, with five-view Test-Time Augmentation (TTA) applied at inference to improve prediction stability. While standalone CNNs achieved around 52% accuracy and ~0.60 AUC, the fusion approach significantly increased these metrics. The highest performance was achieved by classifying Xception and EfficientNet- V2S fusion vectors with XGBoost under TTA, reaching 97.16% accuracy and 0.9914 AUC, with false-positive and false-negative rates of 4.6% and 0.9%, respectively. Grad-CAM heatmaps confirmed that the model's attention was focused on clinically relevant periapical regions, while SHAP analyses revealed that radiomic texture heterogeneity and high-level CNN features jointly contributed to the decision-making process. The proposed method demonstrates high accuracy, balanced error rates, and dual-layer explainability (Grad-CAM + SHAP), offering strong potential for enhancing clinical reliability.