Following his plenary session at PathVisions 2026, Rajendra Singh, Professor of Pathology at the University of Pennsylvania, discusses artificial intelligence’s integration into daily pathology practice as an invisible layer of support throughout the diagnostic workflow.
What does AI-supported pathology look like during a typical working day?
For most of my colleagues, pathology still involves glass slides. Digital pathology remains incomplete, even at large academic centers, and clinical AI is often limited to slide quality control and one or two commercial models. Published advances are substantial, but many departments remain far from implementing them in routine practice. Interestingly, smaller pathology groups, where a colleague may not be readily available to review a case, are often more receptive to AI than larger departments.
Where digital workflows are established, the potential becomes clearer. Consider a 1.1 mm melanoma without ulceration that I sign out in under five minutes. AI can check slide focus and staining, retrieve the patient’s history, assemble and prioritize the case, measure tumor depth and margins, complete the report with the appropriate codes, and convert the findings into computable data for registries or research.
None of this replaces my judgment. It removes friction and gives me back attention – the one resource in a working day that cannot be scaled. That is what I mean by invisible intelligence: the most effective AI is often the AI you barely notice. The pathologist remains in control, and although the interface may become invisible, the evidence and audit trail must remain accessible.
Where can AI provide the greatest value across the diagnostic workflow?
AI currently offers its greatest value before and after interpretation. Pre-analytic applications include slide quality control, specimen and block tracking, identification of missing requisition information, monitoring of scanner and staining variability, and case assembly. Post-analytic tools can support report completion, coding, structured data capture, and conversion of free text into information that registries and clinical trials can use.
These applications may attract less attention than diagnostic models, but they address tasks that consume considerable time. Many may also justify their cost through operational savings without requiring a separate reimbursement code, although laboratories must demonstrate these benefits locally.
The analytic phase is scientifically exciting but more difficult to implement. Potential uses include prioritizing regions of interest, detecting organisms, taking measurements, counting mitoses, and predicting molecular status or prognosis from morphology. Some tools are already in clinical use, but each must demonstrate its value through validation, workflow integration, and a viable reimbursement or funding model.
How can case prioritization, image analysis, and diagnostic support improve accuracy and turnaround times?
Case prioritization is one of AI’s clearest yet most underappreciated benefits. A worklist that moves a suspected melanoma or high-grade lesion forward can reduce delays while ensuring that every case still receives a complete assessment. It also helps pathologists direct their attention effectively. However, prioritization must continue to reflect clinical urgency, and the absence of an alert must never justify leaving a case at the bottom of the queue.
Image analysis can improve consistency in measurements, cell counts, searches for small features across large tissue areas, and standardized reporting. I am more cautious about AI-generated differential diagnoses. The tools I find most useful are knowledge bases and evidence-based search systems that quickly retrieve a relevant reference, prior case, or morphologic pattern.
I favor AI that challenges my interpretation using evidence I can inspect rather than simply providing an answer. Its effect should be measured across the complete workflow – from receipt of the case to delivery of an accurate, finalized report – not solely by the speed of image review.
What clinical impact is AI already having on quality control, patient safety, and treatment decisions?
AI’s most significant contribution so far has been to quality control. It can identify an out-of-focus or understained slide before it reaches a pathologist, detect a mismatch between a specimen and its requisition, or flag disagreement between a diagnosis and the reported stage. Automated slide quality control has been demonstrated in laboratories processing tens of thousands of slides. These applications can prevent errors and improve the images used by downstream algorithms.
Prostate cancer represents an important milestone in using AI to inform treatment. A model combining H&E images with clinical information has been validated using samples and outcomes from randomized trials and has received FDA authorization for prognostic use in a defined patient population. Related research has identified patients who may benefit from hormonal therapy. This provides evidence that morphology contains information about prognosis and treatment response beyond what pathologists routinely report.
However, these remain selected applications. The treatment-benefit findings come from retrospective analyses of completed trials rather than prospective trials in which AI guided treatment.
There is also growing interest in AI-enabled companion diagnostics. AstraZeneca and Roche’s investigational TROP2 assay for lung cancer, for example, combines immunohistochemistry with computational image analysis to identify patients who may benefit from a particular treatment. This is an area in which pathology’s role in treatment selection may expand.
How should laboratories validate, monitor, and govern AI tools?
My first question about any AI tool is the same one we ask about a new assay: does it work in this laboratory? That requires testing it with our specimens, section thicknesses, scanners, staining protocols, and patient populations. A strong result from one test set reflects performance on one dataset under specific conditions; it is not a universal measure of reliability. A model may reach the correct result for the wrong reason or fail when exposed to local pre-analytic variation.
Before implementation, laboratories should validate tools using their own material, including difficult cases, and closely examine clinically significant false negatives. They should also evaluate the pathologist and the tool together. The question is not only whether the model is accurate, but whether it improves performance or encourages overreliance.
Evaluation must continue after deployment because scanners change, stains drift, populations shift, and vendors update their software. Laboratories should retain a fixed reference set and rerun it after updates, monitor discrepancies and failed analyses, and review samples of unflagged as well as flagged cases. Changes to scanners, staining protocols, or software should prompt a documented reassessment proportionate to the clinical risk.
Governance must be built into the system. A standing quality board involving pathology, informatics, and administration should oversee performance. Every tool should have a written intended-use statement and clearly defined limitations. The model version and provenance should accompany every AI-generated output, and a named individual should have the authority to suspend a tool when necessary.
Responsibility remains with the pathologist. Clinical judgment cannot be outsourced, and a model error may recur across many cases. The pathologist must sign the diagnosis they believe is correct, although vendors and institutions retain responsibility for providing a reliable system. AI may be invisible within the workflow, but it must remain visible in the audit trail. Independent, shared validation infrastructure is also needed so that every laboratory does not have to solve these challenges alone.
How can AI-enabled workflows turn routine pathology data into useful evidence for clinical research without increasing workloads?
Data should become computable as a by-product of sign-out rather than creating a second task. A melanoma report already contains the diagnosis, tumor depth, ulceration status, and margins. Too often, these data remain in narrative text and must later be extracted manually. If AI converts the findings into standardized, structured data linked to the slide, tissue block, and report version, laboratories can build reusable research resources through routine care.
If slides are digitized and quality-controlled and reports contain structured data, de-identification and cohort assembly for research or industry validation studies could be largely automated within an institutional review board framework. A melanoma cohort could begin with information already verified at sign-out and link it to treatment and follow-up. Relevant measurements and annotations could also be captured during routine review.
However, “no additional workload” is not a realistic promise. Reliable research resources require data linkage, privacy safeguards, governance, and review of uncertain cases. A signed report is not an appropriate reference standard for every research question, and some studies will require separate expert adjudication. The realistic goal is to avoid repeatedly extracting the same information and reserve specialist input for cases that need it.
Pathology data support molecular prediction, companion diagnostics, and drug discovery. By capturing these data effectively, pathologists can move beyond supplying information and take a more active role in shaping the evidence derived from it.
Over the next five years, how do you expect AI to change the pathologist's role and the delivery of diagnostic care?
When residents ask whether AI will take their jobs, my honest answer is that it will take over parts of them – and that is good news. Those parts include counting, measuring, searching, and the clerical aspects of reporting. What remains is more consequential: professional judgment and accountability.
Five years is a short period in medicine, so my expectations are measured. More laboratories will become fully digital, with much of AI’s value coming from less visible workflow applications. A limited number of predictive models will become reimbursable and enter routine practice. Shared validation infrastructure should also begin to reduce duplicated work. Adoption will remain uneven, and smaller laboratories should be able to use validated tools without building the entire supporting infrastructure themselves.
The pathology report will increasingly become the interface between AI and clinical care. It may incorporate predictions about molecular status, prognosis, and treatment response while distinguishing clearly between what was measured, what was inferred, and the strength of the evidence. Some questions that currently require send-out testing may be answered locally using H&E images and clinical information where the application has been appropriately validated.
Training will also need to change. Residents will still require a strong foundation in morphology, but they must learn to recognize when a model is being used outside its intended setting and explain why they agree or disagree with its output.
The central question is whether pathologists will deliver these clinical inferences or simply supply the slides from which others derive them. Pathologists have an opportunity to assume a more central role in precision medicine, although medicolegal frameworks and institutional responsibilities must evolve alongside clinical practice. The future of the profession will be defined not only by what AI can detect, but by how much responsibility pathologists are prepared to accept.
Newsletters
Receive the latest pathologist news, personalities, education, and career development – weekly to your inbox.
