In this study, we aimed to enhance the accuracy of dental treatments, such as implant insertions, by employing deep learning to precisely identify and delineate the maxillary sinus and mandibular canal in panoramic dental X-rays. To achieve this, we constructed a labeled panoramic dental X-ray dataset. Utilizing this dataset, we proposed an optimal model by integrating UNet with a Convolutional Block Attention Module (CBAM) for segmentation tasks. CBAM effectively reduced false positives through attention mechanisms, improving the model’s precision. Additionally, to evaluate the model’s performance across various environments, we performed external validation using an open dataset. To further improve external validation performance, we introduced a novel augmentation technique by modifying CutMix to suit our specific task requirements. The modified CutMix reduced false negatives by leveraging shape bias, enhancing the model’s sensitivity. Our experiments demonstrated that the developed model achieved not only the highest test performance within the constructed dataset, but also robust generalization performance on external validation. This indicates the generalizability of our model across diverse clinical settings.