The REG2025 Challenge - the REport Generation of Pathology using Pan-Asia Giga-pixel WSIs Challenge - is debuting this year at MICCAI 2025. It will take place today (Satellite Day 2) from 10:30 -12:30 in meeting room DCC1-1F-103.
The goal of REG2025 is to enhance the practicality and reliability of pathology report generation models by ensuring they produce clinically meaningful and high-quality content. In addition, this initiative aims to address the limitations of current AI models in reflecting racial and ethnic diversity by utilizing a broader dataset that includes both Pan-Asia and European data. The challenge dataset comprises 10,494 cases collected from six medical centers across five countries—Korea, Japan, India, Turkey, and Germany—contributing to the development of multicultural and multiethnic medical AI technologies.
It is rare in pathology to release such a large-scale, well-curated in-house dataset, and the report-generation task itself has been gaining attention in both academia and industry. As a result, the challenge organizers saw a strong response with 344 registered participants.
We connected with Yumi Lee, a member of the REG2025 organizing committee, to learn more about the challenge and the motivation behind its creation. Yumi is also a student in the combined master's and PhD program at Ewha Womans University in Korea.
Recent advances in vision-language foundation models have opened new possibilities in medical applications, particularly in image captioning, which generates textual descriptions from images. In pathology, producing textual reports from gigapixel-scale whole slide images requires advanced analysis techniques such as slide-level feature extraction to process and interpret vast amounts of visual data.
While automated pathology report generation is a complex task, it has attracted growing interest for its potential to address workforce shortages, improve diagnostic accuracy, and enhance patient care. However, commonly used NLP evaluation metrics such as BLEU, METEOR, and ROUGE are insufficient in the medical domain, where clinical relevance and content accuracy are paramount.
To address these limitations, REG2025 focuses on:
After the challenge concludes, we plan to request code and brief reports from the top-ranking teams in order to perform a detailed analysis. While we have not yet gathered these materials, the current leaderboard shows a highest score of 0.8249, which is a remarkable figure in the field of pathology report generation. This result raises high expectations for the research and practical potential of the submitted approaches.
We are excited to introduce Faisal Mahmood as our keynote speaker, who will be speaking at the opening of our session. Faisal Mahmood is an Associate Professor of Pathology at Harvard Medical School specializing in computational pathology and AI for clinical decision-making. His laboratory leads pioneering research in AI-driven pathology and has published influential papers in Nature, Cell, and Nature Medicine.
After Prof. Mahmood's keynote presentation, we will hear presentations from the top 3 teams who participated in the challenge, followed by our award ceremony and prize distribution. We hope that many members of the MICCAI community can join us today for the first edition of the REG2025 Challenge.
This challenge is not merely a competition for higher scores—it aims to establish AI-based pathology report generation technologies that can be applied in real clinical settings. Following the final round at MICCAI 2025, we plan to make parts of the dataset and evaluation tools publicly available as open-source resources. We also envision evolving this initiative into a regularly updated benchmark to set a standard for AI-driven pathology report generation.
Visit the REG2025 website for more information.