AI is advancing rapidly in pathology, with new FDA clearances, reimbursement pathways, and growing adoption across health systems. But one critical issue has received far less attention: many hospitals are deploying AI tools without confirming they have been validated for the scanners used to generate digital pathology images. That mismatch can lead to clinically meaningful differences in performance.
A new recommendation statement published in AI in Precision Oncology by the Digital Pathology Association (DPA) aims to address this gap. Here, lead author Nathan Silberman highlights the importance of this comprehensive guidance.
Silberman is not an official spokesperson for the DPA, so any forward-looking opinions are his own. However, direct quotations from the published recommendation statement reflect the DPA's guidance and can be attributed accordingly.
Why did the DPA feel this guidance was needed now, and what gap is it trying to fill for pathology laboratories?
The central challenge is no longer whether AI works in pathology, but how laboratories can implement it safely, consistently, and at scale.
Digital pathology and AI are transforming anatomical pathology, but adoption has outpaced the development of implementation standards. As a result, many healthcare organizations are deploying these technologies without consistent guidance on validation, quality assurance, or performance monitoring. This creates risks, including variable performance between institutions, uncertain regulatory and reimbursement pathways, and potential impacts on patient safety.
The goal of these recommendations is not to determine whether AI should be used in pathology, but to provide an evidence-based framework for implementing it safely, reproducibly, and consistently across laboratories.
The guidance also addresses an important equity issue: digital pathology and AI have the potential to expand access to specialist expertise in underserved settings, but only if they are implemented responsibly and reliably.
Ultimately, these recommendations provide laboratories with a practical roadmap for validating and deploying AI in ways that support high-quality, consistent patient care.
Validation is a major focus of the recommendations. Why is thorough validation so important before AI can be used in routine diagnostics?
Studies have shown that AI models can lose accuracy when analyzing images generated by scanners different from those used during training. These are not isolated examples but well-documented cases demonstrating that even high-performing algorithms may behave inconsistently when deployed in new clinical environments.
That is why the recommendations emphasize validating both accuracy and reliability. They call for quality control measures to detect technical failures and recommend that validation requirements be matched to the clinical risk of the application. For example, an AI tool that highlights areas of interest requires a different level of evidence than one that generates diagnostic conclusions independently. Standardized validation provides laboratories with a consistent framework for identifying potential problems before deployment, helping ensure AI can be introduced safely while maintaining high standards of patient care.
The guidance recommends validating scanners and AI algorithms separately. Why does that matter in practice?
Separating scanner validation from AI validation is important for several reasons.
First, digital slide scanners introduce their own source of variability. Because image acquisition is the foundation for both human interpretation and AI analysis, scanner performance must be assessed independently of any algorithm applied to the images.
Second, this approach is consistent with established laboratory practice. The College of American Pathologists (CAP) already recommends validating whole-slide imaging systems independently of how the images will ultimately be used, whether by pathologists or AI tools.
Third, scanner validation and AI validation require different expertise and methodologies. Scanner validation focuses on image quality parameters such as color calibration, focus, resolution, compression, and image consistency. AI validation, by contrast, evaluates an algorithm's performance against a clinical reference standard. Combining these distinct assessments risks overlooking problems in both the imaging system and the algorithm.
Perhaps most importantly, differences between scanners – including hardware, optics, resolution, color calibration, compression, and file formats – can significantly affect AI performance. Studies have shown that algorithms trained on images from one scanner may perform less reliably when applied to images generated by another.
By validating scanners and AI systems separately, laboratories can determine whether any performance issue originates from the imaging platform or the algorithm itself, allowing problems to be identified and addressed more effectively. This distinction has become increasingly important as regulators have recognized the impact of scanner variability on AI performance.
What do these recommendations mean in practical terms for pathology laboratories that are planning to implement AI?
Innovation succeeds only when clinicians can trust new technology in everyday practice.
The recommendations outline several practical steps for laboratories adopting AI. First, implementation should begin with early planning. Successful adoption requires more than purchasing scanners or software – it also depends on investing in the people, workflows, and IT infrastructure needed to support AI in routine practice.
Second, pathologist oversight must remain central. AI is intended to support clinical decision-making, not replace the expertise and judgement of pathologists.
Third, laboratories should validate scanners and AI algorithms separately and establish minimum sample quality requirements, recognizing that AI performance depends on the quality of the images it receives. They should also implement end-to-end quality control processes to identify technical issues before they affect patient results.
Finally, AI tools must be used only for their intended purpose. Evidence shows that clinicians may apply AI beyond its validated use, making institutional safeguards and clear governance essential.
Taken together, these recommendations provide laboratories with a practical framework for implementing AI safely, consistently, and with confidence.
How do you see regulation and standards helping laboratories adopt AI safely while still encouraging innovation?
The goal is not more regulation – it is more predictable, reproducible patient care.
The recommendations show that innovation and regulatory rigor do not have to be at odds. Instead, oversight should be proportionate to risk and supported by the available evidence.
One example is the approach to interscanner concordance. Rather than requiring laboratories to fully revalidate an AI algorithm for every scanner platform, the recommendations support concordance studies as a practical, evidence-based way to demonstrate that an algorithm performs consistently across different scanners. Requiring full analytical revalidation for every device would be resource-intensive, lead to slow adoption, and create unnecessary barriers to innovation without clear improvements in patient safety.
This principle of proportionality underpins the broader framework. Recommendations are graded according to the strength and certainty of the available evidence, allowing regulators, payers, and laboratories to align requirements with the level of clinical risk and evidence. Similarly, distinguishing AI software that directly supports clinical decision-making from software that simply stores, transfers, or displays data helps focus regulatory oversight where it has the greatest impact.
By taking a risk-based, evidence-driven approach, the framework aims to protect patients while providing a practical pathway for the safe and timely adoption of AI in pathology.
Looking ahead, what developments in AI do you think will have the biggest impact on pathology over the next five to ten years?
I see three developments that are likely to have the greatest impact.
First, AI is beginning to identify patterns in pathology slides that are not detectable by the human eye. This is already being applied in prognostic and predictive modelling, where AI combines pathology images with clinical data to better stratify patients by risk, guide treatment decisions, and support patient selection for clinical trials.
Second, AI is showing increasing potential to predict molecular biomarkers directly from routine H&E-stained slides. Studies have demonstrated that these models can infer features such as microsatellite instability (MSI), tumor mutational burden, and specific genetic alterations with clinically relevant accuracy. Some can even generate virtual immunohistochemistry (IHC) images from H&E slides alone.
Third, AI can deliver more consistent biomarker quantification by reducing interobserver variability in assessments such as Ki-67 scoring and immune phenotype classification. Because biomarker scoring often determines treatment eligibility, improving the reproducibility of these measurements could be one of the most immediate and clinically meaningful applications of AI in pathology.
Newsletters
Receive the latest pathologist news, personalities, education, and career development – weekly to your inbox.
