VQ-Seg uses QPM and a dual-branch design with foundation model-guided alignment to enable controllable perturbations for consistency learning, achieving strong performance on two datasets for medical semi-supervised segmentation.

Learnable Weighted Multi‑Scale Logits Mixing