Integrated and Advanced Deep Learning Framework for Accurate, Interpretable, and Scalable Tooth Defect Diagnosis and Segmentation

Abstract

Accurate, interpretable, and scalable diagnostic solutions are warranted due to increasing incidences of tooth defects. Current methods often fail to generalize and lack transparency, particularly across heterogeneous datasets. An integrated Deep Learning Framework for tooth defect diagnosis and segmentation that relies on the recent EfficientNetB7 in combination with Squeeze-and-Excitation (SE) blocks to enhance the overall feature recalibration, enables the incorporation of Vision Transformers (ViT) for global context modeling, and then adds further convolutional neural networks with Grad-CAM for the visual interpretability process. There is also the use of Attention U-Net for accurate lesion segmentation and SHAP (SHapley Additive exPlanations) to quantify pixel-level attributions. This synergy will achieve very high validation accuracy (92.4%), very high segmentation metrics (IoU > 0.82), and interpretability scores (> 9/10) which can potentially take it as a viable solution in clinical settings. Importantly, the combination of these architectural components represents a considerable breakthrough for AI-driven dental diagnostics.