Diabetic retinopathy (DR) is a leading cause of blindness worldwide. Early diagnosis is essential in the treatment of diabetes and can assist in preventing vision impairment. Since manual annotation of medical images is time-consuming, costly, and prone to subjectivity that leads to inconsistent diagnoses, several deep learning segmentation approaches have been proposed to address these challenges. However, these networks often rely on simple loss functions, such as binary cross entropy (BCE), which may not be sophisticated enough to effectively segment lesions such as those present in DR. In this paper, we propose a loss function that incorporates a global segmentation loss, a patch-wise density loss, and a patch-wise edge-aware loss to improve the performance of these networks on the detection and segmentation of hard exudates. Comparing our proposed loss function against the BCE loss on several state-of-the-art networks, our experimental results reveal substantial improvement in network performance achieved by incorporating the patch-wise contrastive loss.

翻译：糖尿病视网膜病变（DR）是全球致盲的主要原因之一。早期诊断对于糖尿病的治疗至关重要，有助于预防视力损伤。由于医学图像的人工标注耗时、成本高昂，且容易因主观性导致诊断不一致，研究者已提出多种深度学习分割方法以应对这些挑战。然而，这些网络通常依赖二元交叉熵（BCE）等简单损失函数，可能不足以有效分割DR中的病变区域。本文提出一种融合全局分割损失、分块密度损失与分块边缘感知损失的损失函数，以提升网络对硬性渗出物的检测与分割性能。将所提损失函数与BCE损失在多个当前最优网络上进行对比，实验结果表明，引入分块对比损失可显著提升网络性能。