Akademik Veri Yönetim Sistemi
NEUROCOMPUTING, cilt.686, 2026 (SCI-Expanded, Scopus)
Diabetic retinopathy (DR) is a leading cause of preventable blindness, and accurate lesion segmentation plays a crucial role in early diagnosis and treatment. In this paper, we propose a Mamba-based dual-decoder adversar ial network (MDDAd-Net) for semi-supervised, multi-scale DR lesion segmentation. The MambaVision encoder includes convolution, transformer, and Mamba blocks to capture both local and global features within a sin gle backbone. A dual-decoder structure ensures consistency between predictions from the full-resolution and down-scaled images, thereby enhancing scale-invariant feature learning. Furthermore, two adversarial discrim inators are employed: the first reduces feature distribution mismatches between labeled and unlabeled data, while the second aligns segmentation predictions across different scales, thereby refining lesion boundaries. We evaluate our method on two publicly available datasets, IDRiD and DDR. Extensive experiments demonstrate that our model consistently outperforms state-of-the-art methods in terms of the Area Under the Precision-Recall Curve (AUPR), Intersection over Union (IoU), and F1-score, while achieving competitive model complexity. Indepth performance analysis further validates the contribution of each component, showing that scale consistency and adversarial learning are complementary and jointly improve performance. On the IDRiD dataset, our ap proach achieves 73.60% mAUPR, 68.51% mF1, and 53.08% mIoU, while on the DDR dataset, it obtains 52.32% mAUPR, 49.58% mF1, and 33.85% mIoU, surpassing the previous state-of-the-art methods on both datasets. Moreover, experiments on additional datasets from different medical imaging domains demonstrate that the proposed framework achieves state-of-the-art performance and exhibits strong domain-agnostic generalization.