GMS-JIGNet: guided multi-scale jigsaw puzzles for self-supervised artificial spot segmentation in fundus photography.

Dust and sensor noise often create artificial spots in fundus photography, and clinicians may occasionally misinterpret them as pathological signs such as microaneurysms. Reliable computer-aided diagnosis depends on accurately identifying and segmenting such artifacts. However, producing pixel-level annotations for these subtle structures remains labor-intensive and challenging to scale. We propose GMS-JIGNet, a self-supervised segmentation framework based on guided multi-scale jigsaw puzzles and contrastive learning, to address this issue. The method learns spatially-aware representations from unlabeled data by solving jigsaw puzzles across multiple resolutions while selectively injecting positional hints for uninformative regions. The downstream segmentation model receives these representations and uses the ViT encoders from the pretext task as fixed feature extractors and a lightweight FPN decoder. Experimental results on a large-scale fundus dataset show that our proposed model achieves state-of-the-art performance across various metrics, including IoU, DICE, and SSIM, even when trained with only a few labeled images. Moreover, we conducted ablation studies to assess how well our architecture performs under different training hyperparameter setups. The results support the effectiveness of guided self-supervised learning in medical image segmentation and suggest its strong potential for clinical use, especially in settings with limited labeled data.