Researchers from Microsoft, Providence Health System, and the University of Washington have developed a new generative AI model, Prov-GigaPath, for cancer diagnosis. This model, which analyzes over a billion tissue sample images from more than 30,000 patients, was unveiled in a study published today in the journal Nature and is already in clinical use.
Prov-GigaPath, now available as an open-access tool, harnesses AI to detect novel relationships and insights in pathology slides that human eyes might miss. “The rich data in pathology slides can, through AI tools like Prov-GigaPath, uncover novel relationships and insights that go beyond what the human eye can discern,” said Carlo Bifulco, chief medical officer of Providence Genomics. Bifulco emphasized the importance of making this model widely accessible to advance global cancer research and diagnostics.
The development of Prov-GigaPath involved OpenAI’s GPT-3.5 generative AI platform to analyze patterns in 1.3 billion pathology image tiles from 171,189 digital whole slides provided by Providence. This represents the largest pre-training effort in whole-slide modeling, leveraging a database significantly larger than other datasets, including The Cancer Genome Atlas.
Whole-slide imaging, which digitizes tumor tissue microscope slides into high-resolution images, has become a staple in digital pathology. However, these gigapixel slides are vastly larger than typical images, posing a challenge for conventional computer vision programs. Microsoft’s GigaPath platform addressed this by segmenting the large-scale images into 256-by-256-pixel tiles, facilitating the identification of patterns across various cancer subtypes.
To evaluate Prov-GigaPath’s performance, researchers established a digital pathology benchmark comprising nine cancer subtyping tasks and 17 analytical tasks. Prov-GigaPath achieved state-of-the-art results in 25 out of 26 tasks, significantly outperforming the second-best model in 18 tasks, according to study authors Hoifung Poon and Naoto Usuyama from Microsoft.
Poon and Usuyama noted that this AI-assisted approach to digital pathology has the potential to enhance patient care and accelerate clinical discoveries, though more research is needed. “Most importantly, we have yet to explore the impact of GigaPath and whole-slide pretraining in many key precision health tasks such as modeling tumor microenvironment and predicting treatment response,” they wrote.
The study, titled “A Whole-Slide Foundation Model for Digital Pathology From Real-World Data,” includes contributions from Hanwen Xu, Jaspreet Bagga, Sheng Zhang, Rajesh Rao, Tristan Naumann, Cliff Wong, Zelalem Gero, Javier González, Yu Gu, Yanbo Xu, Mu Wei, Wenhui Wang, Shuming Ma, Furu Wei, Jianwei Yang, Chunyuan Li, Jianfeng Gao, Jaylen Rosemon, Tucker Bower, Soohee Lee, Roshanthi Weerasinghe, Bill J. Wright, Ari Robicsek, Brian Piening, and Sheng Wang.
