SAM-SPJunc: Self-Prompting for Junction Detection in Retinal Images via Radius-Based Representations

Minasadat Attari, Kannappan Palaniappan, Filiz Bunyak; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025, pp. 615-623

Detecting junctions in the retinal vasculature is vital to analyze topological structures relevant to disease diagnosis and progression. Although deep learning models have achieved high accuracy in medical image segmentation, their decision making remains opaque, limiting their adoption in sensitive clinical applications. In this work, we propose SAM-SPJunc, a SAM-based Self-Prompted Junction Detection architecture where a dedicated decoder first predicts a radius-aware soft mask that encodes potential junction regions. This coarse prediction is then used as a dense prompt to guide a second decoder that acts as a learnable refinement module generating the final junction predictions through regression to a distance transform. By embedding structural prior knowledge in the form of self-generated radius-based prompts, our model improves spatial focus, reduces false positives, and promotes interpretability. This modular design demonstrates that prompting can serve not only as a means of task control, but also as a foundation for more interpretable and structured medical AI systems.