Zhu Dongsheng, Ge Aiming, Chen Xindi, Wang Qiuyang, Wu Jiangbo, Liu Shuo
Academy for Engineering & Technology, Fudan University, Shanghai 200433, China.
School of Information Science and Technology, Fudan University, Shanghai 200433, China.
Diagnostics (Basel). 2023 Jul 17;13(14):2389. doi: 10.3390/diagnostics13142389.
Many researchers have realized the intelligent medical diagnosis of diabetic retinopathy (DR) from fundus images by using deep learning methods, including supervised contrastive learning (SupCon). However, although SupCon brings label information into the calculation of contrastive learning, it does not distinguish between augmented positives and same-label positives. As a result, we propose the concept of Angular Margin and incorporate it into SupCon to address this issue. To demonstrate the effectiveness of our strategy, we tested it on two datasets for the detection and grading of DR. To align with previous work, Accuracy, Precision, Recall, F1, and AUC were selected as evaluation metrics. Moreover, we also chose alignment and uniformity to verify the effect of representation learning and UMAP (Uniform Manifold Approximation and Projection) to visualize fundus image embeddings. In summary, DR detection achieved state-of-the-art results across all metrics, with Accuracy = 98.91, Precision = 98.93, Recall = 98.90, F1 = 98.91, and AUC = 99.80. The grading also attained state-of-the-art results in terms of Accuracy and AUC, which were 85.61 and 93.97, respectively. The experimental results demonstrate that Angular Margin is an excellent intelligent medical diagnostic algorithm, performing well in both DR detection and grading tasks.
许多研究人员已通过使用深度学习方法（包括监督对比学习（SupCon））实现了从眼底图像对糖尿病视网膜病变（DR）进行智能医学诊断。然而，尽管SupCon将标签信息引入对比学习的计算中，但它并未区分增强后的正样本和同标签正样本。因此，我们提出了角度余量的概念，并将其纳入SupCon以解决此问题。为了证明我们策略的有效性，我们在两个用于DR检测和分级的数据集上对其进行了测试。为了与先前的工作保持一致，选择准确率、精确率、召回率、F1值和AUC作为评估指标。此外，我们还选择对齐度和均匀性来验证表示学习的效果，并使用UMAP（均匀流形近似与投影）来可视化眼底图像嵌入。总之，DR检测在所有指标上均取得了领先成果，准确率 = 98.91，精确率 = 98.93，召回率 = 98.90，F1值 = 98.91，AUC = 99.80。分级在准确率和AUC方面也达到了领先水平，分别为85.61和93.97。实验结果表明，角度余量是一种出色的智能医学诊断算法，在DR检测和分级任务中均表现出色。
