How can AI convince pathologists it’s ready for prime time?

Are We Really on the Cusp of a Breakthrough in Pathology? 

The landscape of cancer diagnostics is shifting. Artificial intelligence now demonstrates the ability to perform core tasks traditionally handled by pathologists—from analyzing complex Whole Slide Images (WSI) and rendering diagnoses, to drafting comprehensive reports in the style of leading institutions, complete with highlighted images supporting the interpretation. Is this the pivotal moment where AI steps into the pathologist's role?

To explore this transformative question, the Digital Pathology Association (DPA) blog presents insights from two leading voices in the field:

• Dr. Mark Zarella, Vice Chair of Digital and Computational Pathology at Penn Medicine, acknowledges AI’s significant strides but provides a critical perspective on the long journey ahead to widespread clinical adoption and the advancements still needed from vendors.

• Scott Kilcoyne, a digital pathology industry veteran renowned for spearheading implementations at hundreds of sites globally and collaborating with many major academic hospitals. Scott offers a pragmatic look at AI's current integration capabilities and the immediate future.

Gain understanding by exploring both expert viewpoints on this groundbreaking development.

How can AI convince pathologists it’s ready for prime time?

Digital pathology systems have been commercially available for over two decades. While early systems tended to focus on education and research use, some practices used these systems to support clinical use cases like telepathology and quantitative IHC scoring. Twenty years later, the era of high-throughput slide scanning has ushered in the ability for practices to potentially digitize the majority of slides generated in their practice with fewer scanners and less manual work, enabling the transition to a fully digital workflow. Yet, despite being technically feasible, the vast majority of pathology practices in the United States have elected not to do so. Often cited are high costs, lack of clear guidance on best practices, and too few compelling use cases. However, as AI tools have begun to appear on the commercial market, along with software platforms to better integrate them into the workflow, this may be starting to change.

There is also a risk that it may not. The high costs of digitization and integration are generally continuing to grow. Additional personnel needed to validate systems, manage deployment, and run the operation places a burden on practices already struggling with staffing. The need for local technical expertise, combined with growing demands placed on informaticists and IT departments, can push large projects without a revenue stream or clear value proposition further down the list. AI might be able to add efficiency, but at what cost?

What’s more, very few practical examples of cost savings, efficiency, quality improvement, and improved user experience ever leave the research laboratory. Reports of large-scale prospective clinical deployments are vastly outnumbered by research studies that boast about AI’s ability to perform tasks on well curated data sets and in controlled environments. Replicating these studies independently on real world data sets at times has revealed troubling insights.

Are the pathologists who have not yet ventured down the path of AI resisting progress or are they waiting for it to become more clinically viable?

To create truly clinical-grade AI, vendors and researchers should meet pathology practices where they are. They should acknowledge the reluctance to adopt these technologies and work to address them. Priority should be placed on:

• working with pathologists to identify use cases that are of the highest value to practices;
• developing best practices for AI deployment and safety, including packaging AI tools with evidence-based sample acceptance criteria and ongoing monitoring strategies;
• critically evaluating the performance of their AI models through the lens of “under what circumstances might it be risky to apply this tool?”
• more convincingly demonstrating an AI tool’s ability to generalize to new practices with different laboratories, slide scanners, patient populations, and pathologists prior to embarking on clinical validation;
• making their study designs more transparent, including ensuring ethical principles are followed throughout the development process;
• creating more affordable solutions whenever possible.

Although many of these considerations are not unique to AI, it is still important to acknowledge the ongoing dialog surrounding AI as we approach a critical inflection point that will undoubtedly shape the regulatory landscape and patient attitudes toward AI for a long time. We have to do it well, and there is a lot of work still to be done to make sure that we do.

Mark Zarella

Vice Chair of Digital and Computational Pathology

University of Pennsylvania

For another expert take on AI's role in pathology's future, read Scott Kilcoyne's perspective here.