DWI-Agent: LLM-Powered Agent for Automated DWI Processing and Clinical Interpretation
Abstract
Purpose
Diffusion-weighted MRI (DWI) and its derived parameters, including apparent diffusion coefficient (ADC) and intravoxel incoherent motion (IVIM) metrics, are valuable for assessing tumor response to radiotherapy. However, DWI processing workflows are complex, and clinical interpretation of ADC and IVIM remains challenging. We propose DWI-Agent, which leverages Large Language Models(LLMs) to address the challenges.
Methods
DWI-Agent is built on LangChain (an agent framework), enabling LLMs to dynamically invoke specialized tools. It implements an end-to-end workflow from raw data processing to clinical interpretation, integrating in-house tools for format conversion, DWI distortion correction, denoising, and ADC/IVIM parameter fitting. In addition, DWI-Agent uses retrieval-augmented generation (RAG) to automatically retrieve relevant literature and generate clinical interpretation reports.
Results
DWI-Agent autonomously and seamlessly orchestrates multiple tools to perform robust DWI processing and interpretation. It uses an AI-driven denoising tool to reduce the standard deviation of homogeneous ROIs by ~60% while preserving signal magnitude; and uses another AI-based tool to improve the mean Dice coefficient between DWI and distortion-free anatomical MRI from 0.648 to 0.955, demonstrating effective distortion correction. In addition, it uses a self-supervised ADC/IVIM fitting tool that reduces over 50% errors compared to least-squares on AAPM-IVIM-dMRI dataset. The agent's clinical interpretation capability was validated in in-house head-and-neck cancer patients undergoing chemoradiotherapy. Using RAG-retrieved literature, DWI-Agent provided differentiated interpretations across three distinct ADC/IVIM response patterns: (1) a marked perfusion fraction (fp) increase masked a true diffusion (Dt) decline, leading to a poor treatment response assessment despite minimal ADC change; (2) concordant increases in ADC and IVIM parameters were identified as consistent with effective treatment response; and (3) Discordant ADC decrease with Dt increase was attributed to fp reduction, indicating partial response and the need for close follow-up.
Conclusion
DWI-Agent integrates LLMs with specialized DWI processing tools, achieving one-stop analysis from data processing to clinical interpretation.