Brian Anderson, MD, Coalition for Health AI (CHAI), discusses opportunities and challenges associated with integrating artificial intelligence (AI) into pathology and oncology practice. He emphasized opportunities in diagnostic support and clinical decision-making, alongside challenges with data quality and infrastructure.

Dr Anderson gave the keynote presentation on this topic at the 2025 College of American Pathologists (CAP) Meeting in Orlando, Florida. 

Transcript:

I'm Brian Anderson, I'm the president and CEO of the Coalition for Health AI, or CHAI.

Can you give a brief description of CHAI and its mission?

CHAI was started about 4 and a half years ago as a private sector-led effort in the middle of the pandemic by a number of doctors that wanted to pull together organizations to develop technically-specific best practices around what responsible use of AI in health looks like. We realized really quickly in some initial conversations that we don't have consensus about what responsible AI is in health and, given how consequential health is in health care delivery, we thought it would be a really good thing to begin building consensus across technology companies, health systems payors, life science companies. And so it's grown from the initial group of 8 that was launched to now include over 3800 organizations – 25% of our members are in the startup space with less than 50 employees.

It's a real cross-section of the health ecosystem, building out what are the industry standards or best practices for the development of AI, the deployment of AI, and the management and monitoring of it, the governance of it in health, that's what we're really excited about and that's what we're committed to. 

What are the most persistent myths you see in discussions about AI in pathology?

There's a lot of excitement in many of the specialties that AI is going to solve a lot of the challenges and a lot of the problems that are facing physicians today. A lot of those problems and a lot of those challenges are not technology related. They're people and processes, and challenges around aligning incentives in the overall structure and delivery of health care in American society. Ao AI is not going to solve a lot of those problems – it's not an “easy button” in that sense. 

Where we are excited and where I do see real promise is AI being used as a tool to provide the kind of clinical decision support or clinical insight that providers previously might've had to go look up somewhere in some remote part of the internet, or some study that they didn't read and that they didn't know about but that has very relevant information. The ability for AI tools to really help with that cognitive decision-making for providers is, I think, where the excitement lies. In the other spaces I'm not too sure that that's kind of where AI is going to show real benefit for pathologists. 

The title of your keynote mentions both “perils” and “promise.” What do you believe is the single biggest opportunity and the single biggest risk for oncologists as AI tools enter pathology practice?

Certainly in the opportunity space, we've seen in many other specialties in health, and candidly not just in the health sector itself, but across various sectors of the US economy, that computer vision algorithms are incredibly powerful and incredibly insightful about identifying, picking up particular kinds of signals that the human eye just isn't as good at doing. We have radiologists, pathologists, various specialties that look at images on screens all the time that are really good and really well-trained, but research time and time again shows that when a provider in one of these specialties where you're looking at images has the added insights and power of an AI tool to call out specific areas on an image or specific kinds of insights about an image may help guide that provider to give a more specific, more timely, faster diagnosis or deeper insight to lead to better clinical care, that's where the real opportunity resides. I think that's where candidly a lot of the excitement is, is that you have computer vision algorithms that can now look at an image and look at the various pixels on the image and be trained on tens of millions of these kinds of images and then render an expert opinion and perspective on what that image might represent or the concern about a particular part of that image and so I think that's the real opportunity space.

I would say kind of building off of that, health care is challenged by an imbalance between supply and demand. We have a lot of demand, we have a lot of patients growing older, getting different diagnoses that happen as you age, like different types of potential cancer or concern for cancer and we have a lack of deeply trained expert specialists across the board, not just in pathology, but certainly in pathology [and] I think one of the real exciting areas is bringing that kind of expertise. A pathologist that is trained, but we only have so many of them, they may be in certain parts of our nation, but not in others. But I think unique to pathology and a specific opportunity with AI is the ability to digitally transmit these images in such a way that if I am a patient that might have a concerning diagnosis, a new diagnosis of a cancer, but it needs to be validated by the biopsy and the slides that are part of that biopsy, if I'm in middle America in a rural environment where it might be that there is no expert pathologist trained in this particular kind of cancer diagnosis now the opportunity exists where those images can be digitally created and sent to a pathologist that might be at a comprehensive cancer center elsewhere, maybe thousands of miles away, that with the assistance of AI, could have those images given an initial diagnosis or perhaps triaged, the images triaged, in such a way that they're flagged for that provider, that pathologist to give urgent, immediate attention to in a way that improves the clinical care for that patient that lives in a rural environment. 

Now, the perils in all of this, and I think we're seeing this again, not just in pathology, but it's about image quality. When these models are developed, they are oftentimes these computer vision algorithms, they're oftentimes trained on the ideal kinds of data sets, and that makes sense. You want to train a robust model that is going to be able to capture all of the important features that the model's going to use to make an insight, a recommendation, etcetera and to do that, you need to have high quality images where the pixels are clear, where they stand out, where the image is not blurred, where the lumens are high. And if you don't have that, if you have an image that is blurred, that is out of focus that isn't prepared in a pristine, exceptional, optimal way that doesn't have the lumens and the brightness that you might need to be able to get those pixels to stand out appropriately or the stain to be performed in an appropriate way, you're going to have suboptimal performance. And I think the reality is, and this is again similar to clinical data, the reality is the real-world experience. The real-world data in this case is the preparation of these slides. The luminosity in creating a digital image behind these slides is not oftentimes optimal, it's not perfect, and so the performance degradation, particularly in computer vision algorithms is a real challenge and the performance degrades. And so one of the real perils is training models on optimal sets of data but then using them in real-world settings where the performance is not going to be as good as we might hope or might want. And so part of the challenge is going to be how do we navigate that process. 

I think an additional challenge is going to be particular to digital pathology, slides [in] high resolution images are large and creating those large high-resolution images requires technical infrastructure to both store, process, transmit, and use and many health systems don't have that technical infrastructure. Those are some of the perils and some of the opportunities. 

How soon do you think AI-driven pathology will start to directly impact diagnostic turnaround times, accuracy, and ultimately patient care?

We're seeing this right now. There are some exciting startups in CHAI that are demonstrating real impact. For example, there's a startup that is working with surgical oncologists to give immediate real time insights into frozen cross sections, and this is an example where a surgeon cuts out a tumor, they want to understand if they have clear margins, so they prepare slides or they send the specimen urgently down to the pathology out of the surgical ward. [The] patient is still waiting there on the surgical bed and the pathologist in an ideal setting is able to read the slide immediately, a frozen cross section, and render an insight and then the surgeon takes the next appropriate steps. If they have clean margins, they can close up. If they don't, then they need to continue exploring. AI tools are now being used in examples like that to generate insight. Yes, it's a clear margin, no, it's not a clear margin really helping the pathologists render that, to your point, a real-time decision that can get back to the surgeon so that you can have an improvement in the patient's care: be that we're not going to close up and we still need to look for tumor and debulking, or we're done and we can close up, we can get the patient out of here and in a more stable situation. 

Similarly, in a less kind of time-urgent space, we're seeing examples frequently of AI tools being used to help inform, give insights to providers about the images that the algorithms that the models are seeing, so it's happening now. I would say certainly, I think like any kind of digital technology and tool, the adoption curve is going to be one where you have early adopters, using it frequently and seeing benefit, and you're going to have later adopters that haven't had that chance yet. I'm excited to see that progression move and see more people start using these tools. 

In oncology specifically, where do you see AI adding the greatest clinical value?

I used to lead a team that was very deep into developing the ontologies and some of the data elements that would be used by AI tools in the future to better leverage that data, clinical data, to make the insights. This effort was called MCODE: Minimum Common Oncology Data Elements. It was in that space that we really, I guess, began to appreciate troublesome and challenging areas of the practice of oncology for providers. Certainly, one of them is in the pathology space, and we've talked about some of those opportunities and excitement to help pathologists in diagnosis of cancer or in identifying clean margins in a debulking of a specific tumor. 

Some of the other examples that I'm really excited about, NCCN guidelines are incredibly complex, with protocols that are challenging to adhere to and challenging for providers to navigate depending upon how a patient is responding and how the tumor is responding and progressing or responding to any particular chemotherapeutic protocol that a patient might be on. AI has the potential to both help the provider and the patient understand how to navigate complex NCCN guidelines and protocols. I am eager to see startup companies, EHR vendors, Empower providers and patients with that kind of information using high quality data about how the tumor's responding, how the patient's responding to offer insights; be it an alternative chemotherapeutic that might be more advantageous and more helpful, a new report that might have come out offering insights to an alternative that the provider or the patient might not be aware of. I think these are examples in terms of the navigation of complex protocols and guidelines that I see.

I think one of the real visceral experiences I think probably many of your listeners have had is individuals that have a new diagnosis with a cancer. Oftentimes it comes with genomic information, either about the tumor or about the individual, and in many of these instances that is very complex, detailed information and providers oftentimes are challenged to keep up-to-date with all of the manuscripts and publications that are coming out — some of those might be about rare mutagenic unique attributes that could inform a patient to potentially be on a specific chemotherapeutic agent that the provider doesn't know about. Those patients now are going to a lot of the AI tools that are available to consumers from many of the big developers like OpenAI or Philanthropic or Microsoft or Google, and putting in their information, their cancer journey, their story, with all of that complex genomic information and then sharing it and getting a real insight. I have personal friends who lived longer lives because they did this,. And then they went and informed their oncologist what ChatGPT heard, and the oncologist realized, wow, I didn't know about that study, or I didn't know about this unique rare mutation, and how it informed the opportunity to be on a new drug or in a new trial that I wasn't even aware of, as saving people's lives. So, those are some of the real exciting spaces in the cancer space that I'm looking forward to in addition to the pathology space.

Source: 

Anderson B. The perils and promise of AI and digital pathology — Myths, realities, and what it all means for pathology practice. Presented at CAP25. September 13-16, 2025; Orlando, FL. Keynote Address.