A Residual Cross-Attention Fusion Network with Adaptive Focal-Margin Loss for Robust Dental Caries Detection from Intraoral Radiographs

In recent years, Automated dental caries detection using intraoral radiographs has attained increasing potential to assist dentists in early diagnosis and treatment planning. Traditional research primarily relies on general-purpose Convolutional Neural Networks (CNNs) such as VGG16 or ResNet, which have demonstrated reasonable performance in medical imaging tasks. The major challenge remains in accurately identifying carious lesions in the occurrence of anatomical noise, restorations and poor image contrast, which often leads to high false positive rates. This research proposes a novel hybrid architecture called ResXformer to address these limitations. Initially, the data is collected from diverse clinical sources and pre-processed to enhance the poor contrast. A depth-wise separable CNN backbone extracts fine-grained features with reduced computational cost. Later, a residual cross-attention fusion mechanism allows bidirectional information flow between the CNN and Transformer branches, enhancing spatial and contextual learning. Further, an adaptive focal-margin loss is introduced that penalizes ambiguous predictions based on per-sample logit variance, reducing false positives near restorations. Together, these steps create a robust, lightweight model tailored for accurate and interpretable dental caries detection in clinical practice.