Agent-Conditioned Multi-contrast MRI Super-Resolution for Cross-Subject

Xinrong Hu, Chao Fang, Yu Chen, Kai Yang, Chun-Mei Feng, Ping Li

Abstract: Multi-contrast MRI super-resolution (SR) techniques require the simultaneous acquisition of multiple contrasts from the same subject, which is often challenging in real-world clinical settings. In this paper, we propose a novel agent-conditioned multi-contrast MRI SR with cross-subject adaptation, termed AgentMRI. AgentMRI is the first attempt to improve the quality of target contrast images using external auxiliary contrasts from different subjects. It expands the traditional attention mechanism from a triplet to a quadruplet format (Query, Agent, Key, Value), where the agent can be trained to capture commonalities from the auxiliary contrast. These commonalities represent foundational anatomical and tissue structure features that are shareable, rather than details specific to a particular contrast. By interacting the agent with the target contrast, AgentMRI dynamically adjusts the model adapting the agent’s knowledge to the target contrast image. This adapting process assists in identifying inherent connections between the auxiliary and target contrasts, even when they are not directly paired. Our extensive testing on fastMRI and clinical datasets demonstrates that our AgentMRI sets a new benchmark, surpassing state-of-the-art methods across various evaluation metrics.