MAMBADENTAL: A BIDIRECTIONAL MAMBA FRAMEWORK WITH CROSS-SCALE FUSION AND GRAPH ATTENTION FOR PANORAMIC DENTAL RADIOGRAPH CLASSIFICATION

Abstract

Panoramic radiography is the most widely used diagnostic imaging modality in dentistry. It provides a comprehensive view of the full dental arch in a single acquisition. Convolutional neural networks and ViTs have shown many advances for automated dental condition recognition. However, quadratic computational complexity, single-resolution feature modeling, and the inability to encode spatial relationships between adjacent teeth limit these existing approaches. Selective State Space Models (Mamba) have linear sequence modeling through input-dependent state transitions, yet their application to dental radiographic analysis remains unexplored.  This study presents MambaDental, a novel architecture integrating Selective State Space Models (Mamba SSM) with three components: Dual-Path Bidirectional Scanning, Cross-Scale Fusion, and Inter-Tooth Graph Attention, for automated multi-class classification of dental conditions from panoramic radiographs.  A dataset of 4,764 panoramic dental X-ray images across four categories (Fillings, Cavity, Implant, and Impacted Tooth) was processed through multi-scale patch embedding. The Dual-Path Mamba SSM processes each scale's patch sequence in both forward and backward directions with a learned gating mechanism. Cross-Scale Fusion combines representations across resolutions via cross-attention. Inter-Tooth Graph Attention models spatial relationships between dental regions as a graph, enabling explicit relational reasoning. Three classifiers, Random Forest, SVM, and Decision Tree, were evaluated with and without preprocessing. The Mamba+RF model with preprocessing achieved the highest performance: 93.6% accuracy, 0.993 ROC-AUC, and 0.935 F1-score. The ablation study confirmed that each component provides incremental gains. MambaDental demonstrates that Mamba SSM-based architectures with bidirectional scanning and explicit spatial reasoning outperform both standard CNNs and attention-enhanced models for dental radiographic analysis. The linear computational complexity of SSMs offers scalability advantages for high-resolution clinical imaging