Diabetic Retinopathy (DR) has become a leading cause of visual impairment and blindness globally, where early diagnosis and assessment of disease severity are important for providing timely clinical treatment. Traditional deep learning-based DR classification approaches commonly demand large amounts of annotated data, which is costly and time-consuming to collect in medical image processing applications. In order to overcome this challenge, this paper presents the framework of Few-Shot Learning using a Relation Network for Diabetic Retinopathy Detection and Severity Grading (RN-DRC), which can effectively learn discriminative features from a few labelled samples and generalize to unseen retinal images. Specifically, in our proposed framework, we integrate VGG16 as the feature embedding component while utilizing ResNet18 as the relation module. The proposed model is tested on the APTOS 2019 dataset for both binary DR Detection and multiclass DR severity grading tasks. The proposed model achieved an accuracy of 99.64% with an AUC of 0.9982 for binary DR detection, and an accuracy of 96.92% with an AUC of 0.9729 for multiclass DR severity grading, demonstrating its effectiveness in distinguishing subtle retinal disease severity levels. Moreover, we conduct ablation studies to evaluate the impact of the number of training episodes on the stability of the model’s performance.