PARDA: A Perceive-Act-Reflect-Distill Agent for Source-Free Unsupervised Domain Adaptation

Source-Free Unsupervised Domain Adaptation (SFUDA) has emerged as a critical paradigm for addressing severe domain shifts in medical image segmentation under strict data privacy constraints, where source data are inaccessible. However, existing SFUDA methods primarily rely on static, open-loop adaptation pipelines, lacking the ability to actively perceive target domain characteristics and use reflection to correct accumulated errors. Meanwhile, Vision Foundation Models (VFMs) and Vision Language Models (VLMs) exhibit strong generalization across modalities and anatomies, yet their high computational inference overhead and deployment constraints prevent direct adoption on clinical edge devices. In this work, we propose PARDA, a VLM-coupled closed-loop SFUDA framework for medical image segmentation. PARDA reformulates SFUDA as a closed-loop decision-making process, in which the method actively perceives target domain characteristics, acts by coordinating semantic and geometric VFMs to generate pseudo-labels, reflects on the reliability of its anatomical and structural representations to guide refinement, and finally distills the acquired adaptation knowledge into a lightweight student model. This design enables PARDA to mitigate error accumulation while preserving data privacy and inference efficiency. We conduct a comprehensive SFUDA evaluation covering 13 cross-modality domain adaptation directions and 24 anatomical targets across diverse organs and imaging modalities. Experimental results demonstrate that PARDA achieves robust and universal domain adaptation while distilling the adaptation results into a lightweight student model for resource-constrained inference.