Supervised sigmoid contrastive loss driven multi-lesion segmentation in fundus images

Automated segmentation of diabetic retinopathy (DR) lesions provides critical decision support for ophthalmologists by enabling quantitative analysis of pathological biomarkers. The lesions often exhibit high inter-class similarity, large intra-class differences and substantial size heterogeneity, posing a formidable challenge for multi-class DR lesion segmentation. In this paper, a Sigmoid Contrastive Loss is proposed that drives the model to learn a discriminative feature embedding space, where features of identical lesions converge while those of distinct lesions diverge. Its core idea is to enforce anchor samples (misclassified pixels) to align closely with positive samples while repelling negative ones, optimizing a structured feature space. To maximize guidance from accurately classified pixels, a segmentation-aware positive and negative sampling strategy with memory bank is proposed by utilizing explicitly semantic information to explore more representative and discriminative features. Moreover, a Coordinate Attention-guided M2MRF (CA-M2MRF) is proposed to capture long-range spatial dependencies between vascular structures and pathological lesions, thereby implicitly exploiting their pathological associations for precise multi-lesion segmentation. The experimental results demonstrate that our method achieved state-of-the-art performance, reaching to 68.32% F1-score, 52.74% IoU and 69.50% AUPR for DR lesion segmentation on the IDRID dataset, and 46.63% F1-score, 31.36% IoU and 49.40% AUPR on the DDR dataset. Extensive experimental results demonstrate the proposed method overcomes the interference of similar tissues/lesions and noises to realize accurate multi-lesion segmentation.