Paige, a global leader in end-to-end digital pathology solutions and clinical AI, today announced the release of Pathology Report and Image Summarization Model (PRISM)1, a whole-slide-level Foundation Model designed to accelerate breakthroughs in clinical decision support and precision medicine.
PRISM builds upon Paige’s landmark development of Virchow2, the first and largest million-slide scale Foundation Model for pathology and oncology. Most recent foundation models in pathology, like Virchow3, focus on small image sections or tiles to make predictions. However, there’s a gap between these tile images and the whole-slide predictions made in medical practice, and as a result, large datasets are still required for AI models to effectively make predictions on whole slides. Additionally, many real-world applications such as for clinical trials with only a few hundred patients, lack enough high-quality curated data to empower accurate AI training.
PRISM is designed to bridge this gap. The new model automatically generates diagnostic summary reports from H&E-stained whole-slide images and can be adapted to additional, advanced downstream tasks in both clinical and pharmaceutical settings. This is a testament to Paige’s innovation and a leap forward in pathology technology towards building a general AI copilot that has the potential to accelerate precision diagnosis and treatment.
Pre-trained on a large-scale dataset of 587,000 whole-slide images and 195,000 clinical reports, the model outputs whole-slide summary reports complete with rich information such as cancer detection, subtyping and biomarker predictions1. By adding this powerful insight to diagnosis, PRISM could directly support reduced diagnostic turnaround times and improve the personalization of cancer diagnosis.
The model also holds great potential for drug discovery and precision medicine. Fine-tuned on biomarkers previously unseen by the model, PRISM was shown to outperform supervised baselines trained on these biomarkers from scratch, especially with low training sample sizes. Notably, PRISM identified certain biomarkers using only 10% of the training data required to meet or exceed the maximum performance achieved without pre-training.1 As biomarker data can be limited, this could help pharmaceutical and life sciences companies to unlock key nuanced information from tissue samples that would refine the way cancer is understood and inform targeted treatments.
“PRISM’s ability to accurately summarize the whole-slides into reports by identifying common and rare cancers along with biomarkers and other deep tissue insight holds extraordinary promise across the landscape of cancer diagnosis and life sciences. Paige, collaborating with Microsoft, is extremely proud to be at the forefront of advancing the Foundation Model development in pathology and oncology with large-scale multi-modal vision and language generative AI. We believe PRISM will help to drive a more detailed look into the complex world of cancer, enhancing diagnosis and treatment for the better,” said Siqi Liu, Director of AI Science.
Read the full publication
—
1Shaikovski, G., et al. (2024). PRISM: A Multi-Modal Generative Foundation Model for Slide-Level Histopathology. arXiv. https://arxiv.org/abs/2405.10254
3Vorontsov, E., et al. (2023). Virchow: A Million-Slide Digital Pathology Foundation Model. arXiv. https://arxiv.org/abs/2309.07778
