Diabetic retinopathy (DR) is a leading global cause of vision impairment. The International Clinical Diabetic Retinopathy (ICDR) system has been proposed to standardize the grading of DR severity . Ultra-widefield (UWF) imaging, which offers a 200-degree view, enhances the detection of peripheral pathology often overlooked by standard imaging methods. Recent studies have demonstrated the effectiveness of deep learning algorithms in analyzing UWF images for automated DR detection, thereby improving sensitivity, specificity, and risk stratification. These advancements underscore the potential of combining UWF imaging with deep learning to enhance DR detection and management, providing a more comprehensive view of disease progression and reducing the rates of ungradable images for more reliable diagnostics. This synthesis highlights the importance of integrating advanced imaging techniques with AI for early and accurate DR detection, which is crucial for preventing vision loss in diabetic patients. To address this, we participated in the Ultra-Widefield Diabetic Retinopathy Challenge, developing algorithms for DR severity and DME classification based on UWF images. We employed EfficientNet-B7, known for its efficiency and generalization capabilities, and enhanced its performance through data augmentation and pseudo-label generation. These techniques expanded training datasets, improved model robustness, and leveraged unlabeled data, resulting in significant performance gains. Training on a 5-fold cross-validation dataset, our optimal model achieved an AUC increase of 0.66%-2%. Our findings underscore the potential of UWF imaging and machine learning to enhance DR management, paving the way for automated systems that streamline DR assessment and timely interventions, ultimately preventing vision loss in diabetic patients.