Numerous studies have emphasized the essential function of video laryngoscopes in tracheal intubation, where providing clear visualizations of throat anatomy significantly raises success rates. Clinically, less-experienced doctors may encounter difficulties due to limited familiarity with these structures, and in emergency situations, achieving intubation on the first attempt is crucial for patient safety. Although deep learning is widely applied in medical imaging for its strong recognition capabilities, there remains a gap in applying it to throat image recognition. In this study, we gathered and labeled datasets from video laryngoscopes, verified by clinical experts. We designed a deep learning segmentation model, MP-UNet, that incorporates a multi-scale feature extraction block and a Pyramid Fusion Attention block. With these modules, the model is particularly effective at managing the feature extraction and fusion requirements for throat structures across different scales. Our model showed a minimum 10% improvement in IoU, the most critical metric, over the original U-Net. Compared to other models, MP-UNet also demonstrated strong segmentation performance on the throat dataset.