Clinical Application of Large Language Models in Generating Pathologic Images.

Area of Science:

• Digital Pathology
• Artificial Intelligence in Medicine
• Medical Education Technology

Background:

• Prostate cancer (PCa) diagnosis relies on accurate histopathology.
• Generating diverse and representative training datasets is crucial for medical education.
• Current resources for pathology training may lack sufficient variety in Gleason grades.

Purpose of the Study:

• To evaluate the capability of DALL·E 3 to create synthetic prostate cancer (PCa) images.
• To assess the realism and accuracy of AI-generated images across different Gleason grades.
• To explore the utility of synthetic images for enhancing pathology education and research.

Main Methods:

• DALL·E 3 was used to generate 30 synthetic PCa images across various Gleason grades.
• Images were created based on standard Gleason pattern descriptions.
• Nine uropathologists assessed image realism and accuracy against actual H&E-stained slides.

Main Results:

• AI-generated images received average realism and representativeness scores of 6.04 and 6.17, respectively.
• Significant variations in scores were observed across Gleason patterns (P < .05).
• Gleason 5 images scored highest, accurately reflecting key pathologic features, though fine nuclear detail was limited.

Conclusions:

• DALL·E 3 demonstrates potential for generating customized pathology images to expand educational resources.
• Ethical considerations, including data falsification risks, necessitate responsible AI implementation.
• Collaboration between AI developers and pathologists is vital for ethical integration into pathology.