An explainable weakly supervised attention-enhanced deep learning model for molar occlusion diagnosis

Accurate classification of molar occlusion plays a paramount role in orthodontic diagnosis and treatment planning. Traditional approaches often rely on manual assessment of intraoral images, which is time-consuming and subject to inter-observer variability. Moreover, existing artificial intelligence-based approaches frequently necessitate anatomical segmentation, which increases the methodological complexity. Therefore, we propose a new training paradigm that combines DenseNet201 backbone with multi-scale convolutional block, feature refinement module, and weakly supervised data augmentation network (WSDAN) for multi-class molar occlusion classification. The feature refinement module integrates squeeze-and-excitation, triplet attention, and global context mechanisms for the extraction of context-aware features. In parallel, a set of handcrafted features, including color statistics, spatial gradients, and edge information, are extracted and fused with the deep image representations through a dual-input architecture. WSDAN employs attention-guided cropping and dropping strategies derived from learned saliency maps to act as a training control mechanism. This enables the model to focus on discriminative regions without requiring tooth-level annotations. Experiments were conducted on a clinically annotated dataset of 1101 intraoral images, spanning five occlusion classes. The proposed model achieved classification accuracy of 95.43%, which was higher than that of an expert orthodontist. The proposed framework outperformed well-established CNN models. Additionally, it maintained consistent performance across diverse imaging conditions, including variable lighting, angulation, and bilateral occlusal views. Model explainability was achieved by combining Shapley additive explanations-based feature analysis with gradient-weighted class activation mapping spatial visualizations. This work offers an explainable solution for automated occlusion analysis, setting a foundation for future applications in DL-driven orthodontics.