Artificial intelligence (AI) is moving quickly into pathology, but adoption may be getting ahead of understanding, according to Eric Walk, MD, chief medical officer of PathAI. During the Panel of National Pathology Leaders (PNPL) Q3 Digital Pathology & AI Focus Group meeting this past summer, Walk urged pathology leaders to make AI fluency a priority—not to turn pathologists into algorithm developers, but to give them enough knowledge to evaluate whether AI tools are appropriate for specific clinical applications.
Pathologists routinely understand the basic science behind technologies such as immunohistochemistry (IHC), PCR, and next-generation sequencing (NGS). Walk noted that many may not have the same familiarity with neural networks, even as AI becomes increasingly involved in diagnosis, biomarker assessment, workflow, and precision medicine.
Understanding what is behind the AI label
Walk's presentation covered basic AI and machine-learning concepts, applications in clinical pathology and precision medicine, emerging uses of generative AI and vision language models (VLMs), and technical, regulatory, ethical, and medical-legal questions.
He emphasized that "artificial intelligence" encompasses several different approaches, including machine learning, deep learning, and generative AI. Most current pathology applications use deep learning, in which neural networks learn patterns from data and make predictions such as distinguishing tumor from non-tumor tissue or quantifying biomarkers.
He also discussed foundation models, which are large models trained on broad datasets and used as a backbone for more specific applications. These models can be adapted for tasks including tumor detection, H&E diagnosis, PD-L1 scoring, and tissue segmentation.
For pathology leaders, understanding these distinctions can help when evaluating products. The appropriate method depends on the clinical task—a model designed to count PD-L1-positive cells, for example, may not be appropriate for identifying previously unknown morphologic patterns associated with a genetic mutation.
AI must also fit the laboratory workflow
Walk also emphasized that algorithm performance is only one consideration when evaluating pathology AI. Digital pathology depends on scanners, image management systems, laboratory information systems, storage, and interfaces, and AI needs to work within that environment.
During the meeting, one PNPL participant described evaluating prostate cancer AI tools that performed well technically but were difficult to incorporate into daily sign-out. Walk recommended considering the user interface, number of steps required to use a tool, and the broader AI ecosystem supported by the image management system.
Potential applications extend beyond diagnosis. Walk described possible uses for slide quality control, case prioritization, workload distribution, tumor identification, cell counting, and biomarker quantification. AI could also help distribute cases according to complexity rather than simply the number of cases assigned.
AI-assisted precision medicine was another focus of the presentation. Walk discussed biomarkers such as HER2 and PD-L1, where visual assessment can vary among observers. AI can potentially identify tumor cells and quantify individual cells, producing a more quantitative assessment.
Machine learning may also allow researchers to extract information from H&E and IHC slides, including cell numbers, staining intensity, tissue areas, and spatial relationships, that humans cannot reliably measure at scale. These data could be used to investigate previously unrecognized morphologic patterns associated with clinical outcomes.
Generative AI raises new possibilities and questions
Walk also discussed generative AI and VLMs, which can work with both images and text. Potential applications include reviewing pathology images, incorporating clinical history, developing differential diagnoses, suggesting IHC stains, and supporting pathology decision-making.
However, the presentation also highlighted unresolved concerns involving false-positive and false-negative results, training-data limitations, explainability, regulation, patient consent, and medical-legal responsibility.
Walk also addressed pathology education, noting the need to ensure residents develop foundational diagnostic skills rather than becoming dependent on AI tools.
While pathologists do not need to become AI experts, they increasingly need to understand enough about the technology to evaluate its clinical applications. As AI becomes more integrated into pathology, that fluency may become an increasingly important part of implementing the technology in practice.
Note: This news summary was generated by AI based on a published press release from PNPL, followed by a review from human editors.
