In this methodology paper, we address the challenge of optimizing the performance of a detector for diabetic retinopathy (DR) from ultra-widefield fundus images, focusing on balancing model accuracy and latency. This MICCAI challenge prioritized “area under the receiver operating characteristic curve” (AUROC) and central processing unit (CPU) time for the detection problem. Therefore, we explored various approaches to enhance these performance measures. We evaluated several convolutional neural networks (CNNs) and found that EfficientNet-B1 gave a high AUROC value of 0.9055 on the UWF4DR test dataset while maintaining a minimal forward pass time. Our experiments included the deployment of transfer learning and data augmentation, revealing that traditional augmentation techniques such as flipping and rotation were less effective in medical imaging scenarios due to the inherent variability and complexity of fundus images. This paper details our methodology, including the rationale behind our model selection and the impact of augmentation techniques, providing insights into optimizing machine learning models for medical image analysis.