News
Microsoft, Providence Health Unveil Whole-Slide AI Model
A new paper details the development of a new AI model that has the potential to drastically improve cancer diagnostics.
The model, dubbed Prov-GigaPath, is now generally available. Co-developed by researchers from Providence Health Systems, the University of Washington and Microsoft, Prov-GigaPath was trained to perform digital pathology tasks such as six types of cancer subtyping (ovarian, liver, brain, kidney, breast and central nervous system) and gene analysis, among others.
The results were promising. "Prov-GigaPath attained state-of-the-art performance on 25 out of 26 digital pathology tasks, with significant improvement over the next best model on 18 other tasks," according to the paper published Wednesday on Nature.
Prov-GigaPath is particularly useful for whole-slide modeling, which is hard to achieve using current technologies because of the sizes of the images and the high resolutions needed to make them useful. Images are more typically consumed in smaller portions, called tiles, which are easier to manage but can lead to clinicians missing critical relationships between one slide to the next.
To overcome this problem, the researchers used Microsoft's super-massive transformer model, LongNet, and its medical vision transformer, GigaPath, to power Prov-GigaPath.
According to a blog post by Providence discussing the paper, "[T]his revolutionary approach has the potential to transform cancer diagnostics by holistically capturing global patterns across the whole slide, allowing for improved predictions around mutations and effective cancer subtyping."
The dataset grounding Prov-GigaPath is enormous. The model was trained using 171,189 slides obtained from over 30,000 patients, spanning 31 tissue types and segmented into 1.3 billion tiles. According to the researchers, this is as much as 10 times larger than other datasets used to pretrain medical models, including the The Cancer Genome Atlas (TCGA).
"To our knowledge, this is the first whole-slide foundation model for digital pathology with large-scale pretraining on real-world data," Microsoft scientists Hoifung Poon and Naoto Usuyama wrote in a separate blog post.
According to Providence, Prov-GigaPath opens new opportunities in preventative cancer care. Its researchers plan to harness the model's ability to identify potentially harmful genetic mutations "to overcome socioeconomic barriers and disparities that currently limit the access to personalized/precision medicine in oncology."
"In short," it said, "Providence will use it to unlock a deeper understanding of how to diagnose and treat a patient's tumor that isn't available when viewing a slide with the human eye."