BACKGROUND

Uveal melanoma is the most common primary intraocular malignancy in adults, yet radiotherapy decision-making for this disease often remains complex and variable. Although emerging generative AI models have shown promise in synthesizing vast clinical information, few studies have systematically compared their performance against experienced radiation oncologists in this specialized domain. This study examined the comparative accuracy of three leading generative AI models and experienced radiation oncologists in guideline-based clinical decision-making for uveal melanoma.

METHODS

A structured, 20-question examination reflecting standard radiotherapy guidelines was developed. Fifty radiation oncologists, each with 10-15 years of experience, completed an open-book exam following a 15-day standardized review. Meanwhile, Grok 3 (Think), Gemini 2.0 Flash Thinking, and Open ai o1 pro were each tested through 10 independent chat sessions. Twelve recognized experts in uveal melanoma, blinded to the source of each submission, scored all answer sets. Kruskal-Wallis tests with comparisons were conducted to evaluate group-level differences in total and domain-specific performance.

RESULTS

Of the 80 total sets (50 from oncologists, 30 from AI), Open ai o1 pro achieved the highest mean total score (98.0 ± 1.9), followed by oncologists (91.5 ± 3.2), Grok 3 (82.3 ± 2.1), and Gemini 2.0 (74.2 ± 3.4). Statistically significant differences emerged across all domains, with human experts particularly excelling in treatment selection but still trailing Open ai o1 pro overall. Completion time was significantly shorter for the AI models compared with oncologists.

CONCLUSION

These findings suggest that advanced generative AI can exceed expert-level performance in certain aspects of radiotherapy decision-making for uveal melanoma. Although AI may expedite clinical workflows and offer highly accurate guidance, human judgment remains indispensable for nuanced patient care.