Answers are grounded in this paper and cite their sources.
Objective. Accurate retinal vessel segmentation is critical for early diagnosis and monitoring of ocular and systemic diseases. However, deep learning methods are often limited by scarce pixel-level annotations and the inability of standard loss functions to preserve vascular topology, leading to fragmented vessel predictions. Approach. We propose SAM-KD a semi-supervised framework that adapts the Segment Anything Model (SAM) for retinal vessel segmentation through a two-stage teacher-student distillation strategy. A LoRA-adapted SAM ViT-B teacher and a lightweight TinyViT-5M student are first trained on limited labeled data. The frozen teacher then generates soft pseudo-labels for unlabeled images, guiding the student through feature-level and logit-level distillation. To improve structural continuity, we introduce a graph-based connectivity loss (L conn ) that penalizes topological discrepancies based on Betti numbers and skeleton length. Main results. Experiments on three public benchmarks (STARE, CHASEDB1, and HRF) using only 20% labeled data show that SAM-KD outperforms other baselines across both pixel-level metrics (Dice, sensitivity, MCC) and topological metrics (clDice, ∆b 0 , ∆b 1 ). On HRF, the proposed connectivity loss reduces ∆b 1 by 46% compared with the strongest semi-supervised baseline (38.2 vs. 71.5). The distilled TinyViT-5M student (5.4 M parameters) achieves inference in 12 ms per image with 4.9 GB VRAM, providing a 5.8× speedup and 2.9× lower memory usage than the teacher, with only a 1.1-point Dice reduction. Conclusion: SAM-KD provides an effective and annotation-efficient solution for retinal vessel segmentation by combining foundation model adaptation, hierarchical distillation, and topology-aware supervision. Its lightweight design makes it suitable for real-time clinical deployment in resource-constrained environments.
No takes yet. Share an insight, caveat, or question.
Duc Tri Phan (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: