Digital pathology can make tissue slides available for remote review, consultation and computational analysis. Yet an AI result is only as useful as the image and workflow behind it. Laboratories need to ask two separate questions: is the digital slide system suitable for diagnosis, and does the AI model perform reliably in this laboratory’s intended use?
Validate the image before the model
A whole-slide image depends on specimen preparation, scanning, display and retrieval. The College of American Pathologists’ whole-slide imaging guideline calls for laboratories to establish diagnostic concordance with light microscopy before using a digital system for its intended diagnostic applications. Its guidance concerns the imaging system; it does not, by itself, establish that a particular AI model is clinically valid.
Test where the model will be used
AI performance reported on one institution’s slides may not carry over to another’s staining methods, scanners, specimen mix or case prevalence. A 2025 systematic scoping review of lung-cancer pathology models identified limited robust external validation as one barrier to clinical adoption. Local evaluation should therefore use representative cases and predefine what counts as a useful result, a missed finding and an unacceptable error.
Design the pathologist’s review
Decide whether AI will prioritize a case, mark an area of interest or provide a measurement. Show its output alongside the original slide, with enough context to inspect and disagree. The pathologist should retain a clear way to correct or dismiss it. When the algorithm or scanner changes, document the version and decide whether additional testing is needed.
Monitor the whole service
After launch, track failed scans, missing tissue, processing delays, disagreements and performance across relevant specimen groups. The FDA’s technical performance guidance highlights that whole-slide imaging devices have technical characteristics and limitations that must be understood for their intended use. Clinical monitoring should also include the effect on pathologists’ workload and turnaround time, rather than relying on model accuracy alone.
These steps are an editorial framework for evaluation, not a claim that any particular AI tool is approved or suitable for a specific diagnosis. Laboratory leaders should apply the rules and professional guidance relevant to their jurisdiction and use case.