Artificial Intelligence for Computational and Digital Pathology

In a two-part series, Eric Walk, M.D., chief medical officer at PathAI, breaks down the different types of artificial intelligence and why pathologists should care about them in this text and video question-and-answer session with Roche Diagnostics. The second interview specifically explains how AI impacts pathology and explores what the future could hold. See Roche’s on-demand video with Dr. Walk.

You talk to pathologists working in all kinds of settings every day. What are the most common questions pathologists ask you about AI? 

It really depends on where the pathologist and their organization are in their digital pathology journey. For those labs earlier in the journey, they are primarily focused on building their digital pathology infrastructure: scanner acquisition, Laboratory Information System (LIS) integration, image storage strategy, etc. For these labs, the tendency is to think about AI as something to do during the next phase, but I always encourage them to think about AI from the beginning because decisions they make now, like the type of image management system (IMS) they choose, can have consequences later. For labs that are further along their digital pathology journey, common questions involve topics, such as AI implementation and integration, ROI, regulatory status, workflow efficiency, processing time and cost.

Why should we care about AI in the path lab?

Just like in many other fields, AI has the potential to be transformative for pathology, bringing value to the entire end-to-end workflow in the AP lab. This technology is unique in that it can pre-process pathology image data prior to it being viewed by laboratory staff and pathologists, generating valuable insights along the way. 

For histotechnicians this could mean automated QC of histology slides, dramatically reducing the need for human review. For lab directors, it could mean AI-enabled laboratory workflow analytics and case distribution based on the complexity and a preliminary diagnosis. For pathologists, it could mean AI-based case prioritization and AI-assisted diagnosis and reporting. This all promises to deliver more accurate and timely information to clinicians and patients, potentially improving lab workflows, precision medicine and health outcomes.

Can you explain AI at its most basic level? 

AI, artificial intelligence, is the concept of using machines to recreate human-level intelligence. Since the field of pathology is image based, most AI pathology applications utilize neural networks as the “machine.” These are multi-layered networks of mathematical “neurons” trained and constructed to output a prediction based on input data, such as tumor detection, mitosis counting, or a biomarker score. In the near future, generative AI approaches, like visual language models, will serve as pathologist assistants, auto-generating reports and making diagnostic and testing recommendations based on the image under consideration.

Tell us about how machine learning and foundation models are used in pathology.

Video segment 10:23-12:37 (2:14)

Pathology advancements in AI heavily rely on machine learning to analyze digital images of tissue samples. Machine learning algorithms are trained using numerous annotated images, where pathologists have identified specific cell and tissue features or slide-level diagnoses. Typically, machine learning models are created from scratch for each problem or task. 

Foundation models (FMs) are large AI “backbone” encoder models that are pre-trained on vast amounts of data, enabling them to serve as the basis for multiple downstream specialized models. FMs in pathology are large collections of embeddings, the numerical representations of the cell, tissue and architectural patterns that pathologists use to diagnose disease. In fact, you can think of FMs as the machine-learning equivalent of pathologist residency training. It is important to understand that FMs are not applications themselves, but instead the FM embeddings are used by specific applications like tumor detection or biomarker scoring.
 

What is AI vs. Artificial General Intelligence (AGI)?

Artificial General Intelligence (AGI) is a concept created in the early 2000s to describe broad artificial intelligence that fully captures human level cognitive skills and abilities including reasoning, planning, and the ability to learn from experience. It was meant as a future state ambition for the field, in contrast to more narrow AI technologies such as IBM’s Deep Blue that beat Gary Kasparov in 1996 and Google’s AlphaGo that beat Go master Lee Se-Dol in 2016. What’s exciting in 2025 is that the field has made tremendous progress, to the point that many experts agree the early stages of AGI have been achieved by the latest large language model chatbots like GPT-4, paving the way for more complete AGI in the near future.

What is Agentic AI?

Video segment 16:11-19:54 (3:43)

Agentic AI refers to the use of AI agents that can perform complex, multi-step tasks as part of larger project goals. Agentic AI goes beyond traditional AI algorithms in that they are able to set their own goals, act independently to pursue those goals, and adjust their behavior in response to changing conditions. Agentic AI is powered by the current generation of large language models that have complex reasoning capabilities. Agentic AI applications in pathology are still being explored, but one can imagine an AI agent pathologist assistant that orders relevant IHC or special stains automatically based on the specific case and drafts the pathology report for editing and finalization by the pathologist.
 

How do the different types of AI match to specific use cases in histopathology?

Different AI methods and architectures are tailored to specific use cases in histopathology. Supervised learning models, such as convolutional neural networks (CNNs) and vision transformers, excel at tasks like quantifying biomarkers such as PD-L1 and predicting gene mutations. Graph neural networks (GNNs) are suited for capturing spatial relationships between cells and tissues, which is crucial in analyzing tumor microenvironments.

Weakly supervised learning, particularly multiple instance learning (MIL), can be used to identify patterns in complex data that are beyond human perception but correlate with endpoints of interest, aiding in tasks like prediction of molecular alteration status. Generative AI, including large language models (LLMs) and vision language models (VLMs), show promise in creating synthetic histopathology images for training data and assisting in report generation, where novel outputs based on existing data are needed.