Novel Contraharmonic Correlative Attention Loss for Microaneurysm Segmentation in Fundus Images

P. Rajith Bhargav

Abstract: In this letter, we propose a novel contraharmonic correlative attention loss (C2AL) to segment microaneurysms (MAs) in fundus images, which are the earliest clinical signs of diabetic retinopathy. Our proposed loss function incorporates the correlation between the attention mask and ground truth to enhance microaneurysms (MAs) segmentation. It effectively combines the binary cross-entropy loss that captures the network's deviation on a global scale with MA-aware local attentive context. The formulation of C2AL prevents the attention module from self-mapping without the need for any additional regularization. Extensive experimental evaluations on benchmark fundus image datasets demonstrate that the proposed C2AL-based MA segmentation is superior than the existing methodologies at very low false positive rate.