Large Language Models for Cardiac MRI Diagnosis Based on Standardized Text Descriptions.

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Imaging

Background:

• Cardiac MRI is crucial for disease evaluation but poses diagnostic challenges, especially in centers with less experienced staff.
• The application of large language models (LLMs) in cardiac MRI diagnostics is currently limited, despite their potential in medical imaging.
• Accurate cardiac MRI interpretation requires specialized expertise, highlighting the need for advanced diagnostic support tools.

Purpose of the Study:

• To evaluate the diagnostic performance of advanced LLMs in cardiac MRI interpretation.
• To compare the diagnostic accuracy of LLMs against human radiologists with varying experience levels.
• To determine if LLMs can provide reliable cardiac MRI diagnoses based on standardized clinical and imaging descriptions.

Main Methods:

• A retrospective study analyzed cardiac MRI data from 805 subjects, including various cardiomyopathies and normal controls.
• Clinical and MRI findings were standardized and inputted into multiple LLMs (GPT-4.5, GPT-4o, Deepseek-V3, Deepseek-R1).
• Diagnostic performance was assessed using frequency-weighted sensitivity and specificity, compared against diagnoses from a medical student and radiologists using McNemar test.

Main Results:

• All evaluated LLMs demonstrated excellent frequency-weighted specificity (0.973-0.983).
• LLM sensitivity was comparable to junior radiologists but superior to medical students, though inferior to senior radiologists.
• GPT-4.5 achieved the highest sensitivity among LLMs, outperforming all except the mid-level and senior radiologists.

Conclusions:

• LLMs can generate accurate cardiac MRI diagnoses from standardized descriptions, demonstrating high specificity.
• The diagnostic performance of LLMs, particularly GPT-4.5, approaches that of experienced radiologists.
• These findings suggest LLMs hold significant potential to assist physicians, especially those with less experience, in cardiac MRI interpretation.