Multi-step Automated Report Generation of Cardiovascular Magnetic Resonance Imaging Based on Visual Large Language Model

The goal of this observational study is to evaluate the accuracy, completeness, and clinical consistency of large language model-generated cardiac magnetic resonance (CMR) imaging reports compared with expert radiologist reports in patients undergoing routine clinical CMR examinations.

The main question(s) it aims to answer are:

Can automatically generated CMR reports produced by a large multimodal model accurately reflect key imaging findings and diagnoses when compared with reports written by experienced cardiovascular radiologists?

How does the quality of generated reports perform in terms of clinical correctness, completeness, and linguistic clarity, as assessed by quantitative metrics and expert review?

If there is a comparison group:

Researchers will compare AI-generated CMR reports with ground-truth reports authored by board-certified cardiovascular radiologists to see if the automated system achieves comparable diagnostic accuracy and report quality across different cardiac pathologies.

Participants will:

Undergo standard-of-care cardiac MRI examinations as part of routine clinical practice.

Have their anonymized CMR image data and corresponding radiologist reports retrospectively collected.

Contribute data that will be used to generate automated CMR reports, which will then be evaluated against expert reports using objective metrics (e.g., diagnostic agreement, entity-level accuracy) and subjective clinical scoring by radiologists.