Intelligent Agent Planning for Optimizing Parallel MRI Reconstruction via A Large Language Model

Parallel magnetic resonance imaging (pMRI) reconstruction needs tedious parameter tuning process for achieving optimal image quality. Although data-driven artificial intelligence (AI) has significantly improved pMRI reconstruction, knowledge-driven AI has been little utilized for optimizing pMRI reconstruction. Recent progress of large language models (LLM) embodying vast knowledge bases are adept at decomposing complex tasks into structured and planned steps in some automation tasks. In this paper, we develop an intelligent agent equipped with LLM-based planning capability for optimizing pMRI reconstruction. Based on the existing empirical knowledge of optimal parameter tuning for GRAPPA reconstruction, Planning Domain Definition Language (PDDL) domain and problem files are generated by using an LLM. Then, structured PDDL is used to guide GRAPPA reconstruction. Experimental results show that LLM-based planning can specify clear goals of parameter tuning from unstructured knowledge description and improve image reconstruction. The proposed method may help users such as MRI technologist who is not familiar with pMRI reconstruction to optimize image quality. Future work may eliminate the need of human intervention for fully automatic reconstruction.