Caries segmentation in panoramic X-ray images is an important prerequisite for early caries detection and subsequent treatment. In order to achieve accurate and automatic segmentation of caries in panoramic X-ray images, a semi-supervised learning framework with multi-scale convolution and selective kernel dual-attention mechanism is proposed. This framework aims to enhance the generalization capability of the model by leveraging a large amount of unlabeled data, while addressing challenges such as blurred lesion boundaries and low contrast in caries-affected regions. The framework adopts a teacher-student dual network architecture. It applies multi-scale convolutional attention to deeply supervise the multilayer decoder in the student network, thereby improving its ability to capture boundary details and distinguish between similar inter-class regions. Meanwhile, a selective kernel attention mechanism is introduced to fuse multi-level predictions from the teacher network, adaptively selecting convolution kernels based on pixel-level uncertainty to generate accurate uncertainty masks that guide the student′s learning process. Experiments conducted on the dataset 1 and dataset 2 demonstrate that, on 265 slices, the dual attention mechanism achieves improvements over the baseline model of 3.91%, 2.14%, and 5.35% in Dice coefficient, precision, and sensitivity, respectively. And on 530 slices, the improvements reach 1.39%, 5.69%, and 12.34%, verifying the method′s stability and adaptability on larger-scale data. Compared with traditional fully supervised models, the proposed method achieves the highest improvements in Dice coefficient, precision, and sensitivity, with increases of 22.27%, 17.64%, and 24.57%, respectively. And compared with recent semi-supervised methods, it achieves improvements of up to 14.54%, 14.81%, and 11.96%, respectively. This study not only significantly enhances caries segmentation performance but also provides an accurate and robust solution for panoramic X-ray images.