Skip to main content
Premium Trial:

Request an Annual Quote

Mayo Clinic Pushes for Digitization in Pathology With Partnerships on Foundation Models

Premium
MayoClinic

NEW YORK – The Mayo Clinic aims to use its bank of 20 million whole-slide images of tissue samples and clinical data to aid the development of artificial intelligence-driven pathology tools.

In an interview, Jim Rogers, CEO of Mayo Clinic Digital Pathology, said that Mayo, AI model developer Aignostics, and Nvidia have been developing foundation models that leverage Mayo's trove of patient slide samples. The foundation models that are produced through the project, he said, could be used by Mayo and its partners, as well as pathology application developers and researchers, to develop tools including AI-driven algorithms for cancer detection.

Last week, the Mayo Clinic announced that it had formed a digital pathology platform to apply its extensive archive of pathology slide images to the development of AI-based applications in pathology and accelerate medical breakthroughs. Nvidia is providing technical expertise and accelerated computing capabilities. That support involves the use of the Nvidia Clara suite of computing platform, software and services for the AI-aided development of healthcare applications.

The collaboration with Aignostics, which has offices in Berlin and New York, has included efforts over the past two months to develop a foundation model from 1.2 million de-identified whole-slide images. Aignostics, a spinout of the university hospital Charité-Universitätsmedizin Berlin, plans to continue working with Mayo on the development of a foundation model that is trained on 5 million slides as well as other models.

Viktor Matyas, Aignostics cofounder and CEO, said that his company has experience developing a foundation model using European data, and it has ambitious targets for further foundation model development in the partnership with Mayo. Mayo's expertise and large datasets can help to ensure that the model is trained on representative data for various diseases including rare diseases, he noted.

Matyas said that Aignostics plans to use the foundation models from the Mayo partnership to develop algorithm-based tests for the detection of cancer, but he also wants to see other diagnostic developers use the model.

The creation of Mayo Clinic Digital Pathology has been in the work for years, and Mayo is working with investors and data providers to continue building the platform. The pathology platform is integrated with the Mayo Clinic Platform global innovation network and Mayo Clinic Laboratories, which provides advanced testing and pathology services to healthcare organizations in the US and globally.

Rogers said that it's not yet clear how the models developed through the Mayo program will be monetized, although he said that the plan is to reinvest revenues to continue to improve the tools that are developed through the project.

He also noted that, in addition to the partnerships with Aignostics and Nvidia, Mayo has a longstanding relationship with Google, and Mayo separately announced last week partnerships with Cerebras Systems and Microsoft Research for generative AI research for the development and testing of foundation models that will be used to aid the analysis of radiology images.

Rogers said that only a small portion of pathology labs have purchased digital pathology scanners and analysis software, and they need a compelling reason to invest into the technologies and move beyond microscopy-based slide examination.

The high cost of converting a lab to digital pathology, which can run in the hundreds of thousands to millions of dollars, is frequently cited by digital pathology experts as a primary barrier to investment into slide scanners, displays, slide image management software, and data storage infrastructure though they also said that new entries to the market could lower equipment prices and increase the capabilities for labs that adopt the technologies, making the case for investment more compelling. Digital pathology has been a trailblazer in the adoption of AI-based technologies.

The digitization of glass slides offers healthcare providers access to a deeper well of patient data, Rogers said, and the proliferation of pathology tools could encourage labs to move beyond microscopy-based slide examination. He noted that pathology labs have a role in every cancer case, and the Mayo Clinic intends through this program to make it easier to develop digital pathology applications that can help clinicians to analyze slide images with greater speed, efficiency, and insight.

Creating those applications will be easier, and it will require less data using foundation models, he said. Foundation models are trained on large datasets, and they can be used as the backbone for various tasks with minimal extra training, such as the development of AI-based algorithms that are used to analyze whole-slide images. Others including researchers at Harvard Medical School and the AI-based cancer test developer Paige have also recently reported progress in the development of foundation models for digital pathology.

Earlier this month, researchers from Mayo, Aignostics, and Charité-Universitätsmedizin Berlin published on the preprint server ArXiv study results on the evaluation of the model, Atlas.

Atlas was trained on 1.2 million slides from 490,000 pathology cases from the Mayo Clinic and Charité-Universitätsmedizin Berlin, and according to the authors, the model performed well across benchmarks for molecular and morphology-related tasks, although integrating more data or increasing the model size could yield improvements.