Microsoft unveils whole-slide model for digital pathology

por Gus Iversen, Editor in Chief | May 29, 2024
Artificial Intelligence Business Affairs Pathology
In partnership with Providence Health System and the University of Washington, Microsoft has announced the first whole-slide foundation model for digital pathology that is pretrained with real-world data.

Digital pathology enables the analysis of high-resolution digital images of tumor tissues. GigaPath, the newly developed vision transformer, aims to overcome the computational challenges posed by gigapixel whole-slide images using dilated attention to manage the large-scale data inherent in digital pathology.

The model, which was trained on over a billion pathology image tiles from 170,000 slides, has shown promise in cancer classification and pathomics tasks, highlighting its potential to enhance clinical research and patient care, according to Microsoft. For instance, it achieved a 90% or higher area under the receiver operating characteristic curve (AUROC) for six cancer types, significantly outperforming existing models in numerous tests.

Additionally, GigaPath excels in pathomics tasks by identifying genetic mutations based on slide images. This capability could reveal subtle morphological-genetic connections, enhancing understanding of cancer biology.

GigaPath's capabilities extend to vision-language tasks by incorporating pathology reports into the analysis. By aligning slide representations with textual reports, GigaPath outperforms other models in tasks like zero-shot cancer subtyping and gene mutation prediction, showcasing its potential for whole-slide vision-language modeling.

The development of GigaPath may mark a significant step toward multimodal generative AI in precision health. Future research will explore its application in key areas such as tumor microenvironment modeling and treatment response prediction.

In addition to GigaPath, Microsoft has other digital pathology research collaborations, including partnerships with Cyted geared toward esophageal cancer detection, Volastra for KIF18A inhibitors, and Paige for developing new clinical applications and computational biomarkers.

You Must Be Logged In To Post A Comment