CariesAttNet: An Attention-Enhanced Encoder-Decoder Network for Automated Caries Segmentation in CBCT Images

Dental caries is a prevalent bacterial infection, and its early and accurate detection is essential for preventing irreversible damage to dental structures. Cone-beam computed tomography (CBCT) offers high spatial resolution and clear anatomical visualization, enabling precise assessment of caries lesions. However, the manual layer-by-layer annotation of caries regions in CBCT images is not only time-consuming and inefficient but also heavily reliant on professional expertise, which can easily lead to inconsistencies in segmentation results. To address this challenge, we propose CariesAttNet, a novel attention-guided convolutional neural network designed for automatic caries segmentation. The network adopts an encoder–decoder architecture with skip connections, featuring an attention ConvNeXt block that enhances local feature extraction, a large kernel attention module that captures long-range contextual information through convolutional attention, and a multi-scale attention gate that effectively aggregates features across multiple semantic levels. Experimental results on a private CBCT dataset demonstrate that CariesAttNet outperforms existing mainstream methods across all evaluation metrics, achieving dice similarity coefficient and intersection over union scores of 80.85% and 72.01%, respectively, on the test set. Further ablation experiments and interpretability analyses validate the effectiveness and synergistic benefits of each module.