DOI: 10.1200/cci-24-00267 ISSN: 2473-4276

Clinical Application of Large Language Models in Generating Pathologic Images

Lingxuan Zhu, Yancheng Lai, Na Ta, Weiming Mou, Rodolfo Montironi, Katrina Collins, Kenneth A. Iczkowski, Fei Chen, Antonio Lopez-Beltran, Rui Zhou, Huang He, Gyan Pareek, Elias Hyams, Dragan Golijanin, Sari Khaleel, Borivoj Golijanin, Kamil Malshy, Alessia Cimadamore, Xiang Ni, Tao Yang, Liang Cheng, Rui Chen

PURPOSE

This study investigates the potential of DALL·E 3, an artificial intelligence (AI) model, to generate synthetic pathologic images of prostate cancer (PCa) at varying Gleason grades. The aim is to enhance medical education and research resources, particularly by providing diverse case studies and valuable teaching tools.

METHODS

This study uses DALL·E 3 to generate 30 synthetic images of PCa across various Gleason grades, guided by standard Gleason pattern descriptions. Nine uropathologists evaluated these images for realism and accuracy compared with actual hematoxylin and eosin (H&E)–stained slides using a scoring system.

RESULTS

The average realism and representativeness scores were 6.04 and 6.17, indicating satisfactory quality. Scores varied significantly among Gleason patterns ( P < .05), with Gleason 5 images achieving the highest scores and accurately depicting critical pathologic characteristics. Limitations included a lack of fine nuclear detail, essential for identifying malignancy, which may affect the images' diagnostic utility.

CONCLUSION

DALL·E 3 shows promise in generating customized pathologic images that can aid in education and resource expansion within pathology. However, ethical concerns, such as the potential misuse of AI-generated images for data falsification, highlight the need for responsible oversight. Collaboration between technology firms and pathologists is essential for the ethical integration of AI in pathology practices.

More from our Archive