An Intelligent Hybrid Deep Learning Framework for Automated Dental Image Analysis and Clinical Decision Support Using Explainable Artificial Intelligence

Anatomical overlap, heterogeneous abnormalities, class imbalance, and prediction interpretability are still challenges in automated dental radiographic analysis. This research proposes a hybrid deep learning model combining EfficientNetV2-S and Swin Transformer-T with Lesion-Aware Gated Attention Fusion (LAGAF) for the detection of dental objects and reporting for clinical decision support in multicategory. It involved multiscale Feature Pyramid Network aggregation, localization with Faster R-CNN, visual attribution, confidence calibration and rule-constrained reporting. For experiments, OralXrays-9 was used, which contains 12,688 panoramic images in 9 annotated categories, of which 2,688 images are held out and used for independent evaluation. Compared to independent CNN and Transformer baselines in terms of COCO detection metrics, fusion ablations, computational benchmarking, and 2,000 paired image-level bootstrap resamples. Hybrid-A2 got the best held out [email protected]:0.95 (0.5843), precision (0.8253), and F1-score (0.8790), while decreasing false-positive detections by 31.2% and 36.4% compared to CNN and Swin-T, respectively. But, there was a decrease in recall of 0.9401, pointing towards lesser detection power. The improvement over CNN was confirmed by bootstrap analysis, while the improvement over Swin-T was not clear. During development, there was a slight improvement in the results of the fixed-gate fusion methods, suggesting that learned gating did not consistently lead to improved performance with added complexity. Without independently establishing attribution faithfulness, calibration improvement and clinical effectiveness, explainability and confidence-aware reporting were implemented. These findings demonstrate the computational feasibility of hybrid dental detection and highlight the need for improvements in pathological sensitivity, external validation and clinician-supervised evaluation to facilitate clinical deployment.