November 29, 2023

How Foundation Models Can Transform Pathology

A Q&A with AI scientists working in field of pathology


Paige was founded with the mission of revolutionizing cancer care. To do so, it is critical that we dedicate ourselves to tireless innovation and pushing the boundaries of what is possible with digital pathology technology.

Leading this charge are Brandon Rothrock, PhD, an expert in computer vision, machine learning, robotics, and autonomous systems, and Siqi Liu, PhD, a leader in industrial scale machine learning and medical data analysis. Together, they have been pioneering our latest work on Foundation Model development.

We sat down with Brandon and Siqi to learn more about what foundation models are, how their development might impact pathology, and what this work means for the future of cancer care:

What is challenging about building pathology AI today?

First, AI systems for detecting cancer require extensive data to accurately generalize across various organs and cancer subtypes, which is difficult to gather and process. Secondly, ensuring the accuracy of the ground truth data, essential for training AI, is challenging at large scales. Last, the subtle and diverse patterns in histopathology, again crucial for cancer detection, demand advanced algorithms for precise recognition. As we consider how to build systems for increasingly uncommon or rare cancers, the burden of sourcing sufficient sample sizes exacerbates these challenges, and quickly becomes infeasible. Lack of data availability to develop models for rare biomarkers or drug response makes the problem even more dire.

What are foundation models and how can they be applied to pathology?

Foundation models are a general term for large-scale models trained on expansive and diverse data. Typically, these models are trained using self-supervision and do not need ground-truth labels, but can be effectively adapted to specific applications using a relatively small volume of labeled data. This presents a direct solution to the data cliff problem.

The foundation model being developed at Paige is a large-scale model trained on the natural distribution of all cases handled at one of the world’s leading cancer centers. This dataset, which is millions of slides in volume, consists of all tissue types and cancer conditions. Although the model is still under development, we have already demonstrated excellent performance on many very diverse tasks.

Are there any challenges with the foundation model approach?

Yes, challenges still exist in developing foundation models for pathology. One of the most prominent challenges is handling extreme data imbalance. The natural prevalence of cancer types follows a long-tail distribution, resulting in rare cancers forming the highest number of cases. Furthermore, only a small number of slides within a case, or potentially only a small foci within a single slide, may contain the cancer of interest. The challenge is to design the AI algorithms and training procedures to learn how to differentiate the histologic patterns of interest, particularly when those patterns are not labeled with ground-truth and may be dominated by an overwhelming amount of confounding patterns.

What is unique about the foundation model Paige is currently developing?

The Paige Virchow model marks a revolutionary advance in AI for cancer detection. It is the first model to be trained on a million-scale dataset, a feat that required substantial compute resources and engineering efforts, setting a new benchmark in the field. Further, the model employed cutting-edge self-supervised learning algorithms, eliminating the need for manually annotated ground truth. This approach allowed for greater scalability and adaptability in training. Specifically tailored for digital pathology, the model also included adjustments to better handle the unique characteristics of digital pathology images, enhancing its effectiveness in this specialized area.

How will Paige’s foundation model bring new capabilities to pathology?

Our foundation models in pathology are currently used to catalyze both the improvements of our existing models and for the development of new diagnostic and biomarker AI algorithms. The ability to rapidly adapt the foundation model to new tasks while maintaining a high level of performance and generalization allows us to build mature models much faster. Paige is also investigating exposing this AI development workflow externally to allow 3rd parties to rapidly develop new AI systems using our foundation model in a manner that preserves data privacy and still retains the advantages of large-scale pre-training.

Beyond building conventional diagnostic and biomarker detection capabilities, there are potentially many novel and exciting applications of a foundation model in pathology that are currently speculative, but have the potential to be transformative. This could include allowing pathologists and scientists to interact with the system in a natural way to explain a prediction, cooperatively discover new insights, or automatically advise or create reports and analytics.

How do we develop these models safely and responsibly?

As with any AI model for healthcare, safety and responsibility are paramount to realizing such benefits. We address the safety and reliability of our foundation model in at least three key ways; First, we do not expose the foundation model directly to the user. Second, we rely on conventional testing and validation for the safety of our AI systems. Lastly is the innate objectivity of the data used for development. Unlike generative AI systems developed in other industries that rely heavily on innately biased human-generated data such as natural language or human feedback, our foundation model learns directly from histology imagery that is free from this influence. As such, we can rely on established and mature methods for dealing with bias and performance qualification.

The current version of the Virchow foundation model Paige has created has already marked a turning point in pathology AI. Now, we are continuing to develop this model to go above and beyond what is possible in pathology today. As we accelerate pathology AI technology further, we hope to unlock novel capabilities that can redefine the way pathologists work and ultimately impact patient care for the better.