Dental caries is a significant public health concern and one of the most common dental diseases worldwide. Early detection of proximal and incipient caries plays a critical role in ensuring timely treatment, preventing extensive restorative procedures. To address limitations in traditional detection methods, this work presents a Deformable Graph Convolutional Network with Swarm Bipolar Algorithm (DeGCN-SBA) for enhanced dental caries detection from bitewing radiographs. The process begins by collecting data from the 4th Nord-Trondelag Health Study (HUNT4) Oral Health Study dataset, which provides comprehensive oral health information, including clinical examinations, bitewing radiographs, and patient demographics, supporting reliable research in caries detection. The collected radiographs undergo pre-processing using the Cubature Kalman Filter (CKF), which effectively removes noise and enhances image quality, improving the accuracy of subsequent analysis. After pre-processing, the Clifford Fourier Mellin Transform (CFMT) extracts robust, transformation-invariant features from dental regions, capturing essential patterns of teeth and carious lesions despite variations in rotation, scaling, and translation. These features are then analysed by the Deformable Graph Convolutional Network (DeGCN), which models relationships between clinical attributes and spatial patterns within radiographs, enabling accurate classification of carious and healthy regions. Finally, the Swarm Bipolar Algorithm (SBA) optimizes the network’s weight parameters, enhancing learning efficiency and improving diagnostic performance. The system is evaluated using accuracy (99.93%), precision (99.82%), recall (99.67%), and F1-score (99.72%), demonstrating significant improvements in reliable and efficient dental caries detection.