糖尿病视网膜病变DR是世界上导致视力下降的主要原因。在过去几年中基于人工智能的方法已被用于检测和分级DR。在传统的监督学习问题中标记一个完整的数据集十分昂贵并且标记难以大量获取。针对一些特定任务只有行业专家才能为样本做准确的标记。而主动学习方法尝试通过选择标记较少的样本训练出较好的模型来降低成本。与现有的主动学习方法侧重于选择信息量最大的样本不同在本文中我们提出了一种基于对比编码的混合选择策略的主

Diabetic retinopathy (DR) is a leading cause of visual impairment worldwide. In recent years, artificial intelligence (AI)-based methods have been used for detecting and grading DR. In traditional supervised learning problems, labeling an entire dataset is expensive and difficult to obtain in large quantities. For certain tasks, only industry experts can provide accurate labeling for samples. Active learning methods attempt to reduce costs by selecting fewer labeled samples to train better models. Unlike existing active learning methods that focus on selecting samples with the highest information value, in this paper, we propose an active learning framework based on contrastive coding and a hybrid selection strategy to choose samples that are both highly uncertain and representative. We found that using contrastive learning to train a pre-task learning model on unlabeled datasets and batch processing samples based on contrastive loss helps us select the most uncertain samples for labeling, achieving a hybrid sample selection strategy based on uncertainty and representativeness. A large number of experimental results demonstrate the effectiveness of our proposed method, which achieved an accuracy of 49.3% on the APTOS2019 diabetic retinopathy dataset, outperforming existing active learning methods.