Digital pathology is an innovative field at the intersection of digital transformation in biomedicine and generative AI technology. Whole-slide imaging, which converts microscope slides of tumor tissue into high-resolution digital images, plays a crucial role in deciphering tumor microenvironments for precision immunotherapy. However, handling gigapixel whole-slide images poses computational challenges that conventional vision transformers struggle with. In response to this, a novel vision transformer called GigaPath has been developed to address these challenges, allowing whole-slide modeling by leveraging dilated self-attention.
In collaboration with Providence Health System and the University of Washington, Prov-GigaPath has been introduced as the first whole-slide pathology foundation model pretrained on real-world data. This model has shown state-of-the-art performance on cancer classification, pathomics tasks, and vision-language tasks. It outperforms previous models in various tasks, demonstrating its potential for advancing patient care and clinical research.
Prov-GigaPath’s performance in cancer subtyping, gene mutation prediction, and other pathomics tasks has been successful, with significant improvements over existing models. It has the capability to differentiate among various cancer subtypes and has achieved high accuracy in predicting genetic mutations based on slide images alone. Furthermore, by incorporating pathology reports, Prov-GigaPath has shown promise in vision-language tasks, outperforming three state-of-the-art pathology vision-language models.
The development of GigaPath and Prov-GigaPath represents a significant step towards multimodal generative AI for precision health. Through ablation studies and exploration of whole-slide pretraining and vision-language modeling, best practices have been established. Future opportunities include exploring GigaPath’s impact on modeling tumor microenvironments and treatment response. Prov-GigaPath’s success underscores Microsoft’s commitment to advancing AI in healthcare, with collaborations in digital pathology research leading to exciting progress in the field.
Overall, Prov-GigaPath’s performance on a variety of tasks highlights its potential to revolutionize healthcare practices and accelerate clinical discovery by leveraging the power of AI and digital pathology. The collaborative efforts with various institutions and teams within Microsoft signify a dedicated commitment to advancing AI technology for precision health.
Article Source
https://www.microsoft.com/en-us/research/blog/gigapath-whole-slide-foundation-model-for-digital-pathology/