The Proof Problem, Part 1: AI Isn’t Replacing Our Job. It’s Adding a New One.

Every few months, someone tells me AI is coming for pathology. More recently, the message has become reassuring: AI isn’t coming for our jobs, it’s just coming for the boring parts. I don’t think either framing gets the real change quite right.

A recent talk by Princeton computer scientist Arvind Narayanan gave me a different way to think about it. His argument, in short, is that as AI gets better at doing things, human work doesn’t disappear. It moves from building and executing to evaluating, deciding, and being accountable for the outcome. I think that distinction matters enormously in medicine. Narayanan uses the analogy of a crane operator: a more powerful crane doesn’t eliminate the operator; it changes the job from doing the lifting to directing it and being responsible for what happens. In some ways, that may be what is beginning to happen to us.

Think about what is already happening. A pathologist can experiment with a classifier far more easily than a few years ago. A radiologist can receive AI-generated findings. A cardiologist can get an automated ECG interpretation. Soon, almost every specialty will have models sitting somewhere between clinical data and clinical decisions. Building the model is getting easier. Proving that it deserves to influence patient care is not. That may turn out to be one of the most important responsibilities physicians have in the AI era.

Someone still has to ask: Does this model work on our patients? Does it fail differently in different populations? What happens when the scanner, stain, laboratory, protocol, or patient population changes? Is it learning the biology we think it is learning, or exploiting a shortcut we haven’t noticed? And perhaps most importantly, how much should I trust it when there is a real patient on the other side of the prediction? AI can help us do parts of that evaluation, but the standard for what counts as enough evidence, and the accountability for acting on it cannot simply be delegated back to the system being evaluated. That requires clinical judgment.

For years, much of the conversation has been about whether physicians need to learn to build AI. I increasingly think that is only half the question. We also need physicians who know how to validate AI and to determine when, where, and how much it can be trusted. That doesn’t mean physicians shouldn’t learn to build models, or that model development belongs only to engineers and vendors. Quite the opposite: understanding how these systems are built will make us better users and better evaluators of them.

But building and validating are different responsibilities. Physicians have always had to evaluate evidence before acting on it. AI doesn’t create that responsibility from scratch; it expands it. Determining whether an AI system is safe, generalizable, clinically meaningful, and trustworthy is not only a technical question. It is a clinical one too.

And right now, there is a problem: in most health systems, we have not built the infrastructure to do this well. The vendor has evidence. The paper has an AUC. The regulator may have cleared the product. IT can deploy it. But who determines whether it works on our patients, in our workflow, with our instruments, and under the conditions in which we will actually use it?

That is the proof problem, and I suspect solving it will become one of the defining responsibilities of medicine in the AI era.

Over the next few posts, I want to get practical about what that actually means: what to ask an AI vendor, what makes a validation convincing versus what just sounds convincing, how models quietly fail after they’re deployed, and what infrastructure health systems will need before AI can be trusted as part of routine clinical care. I’ll use pathology as my laboratory, because that’s where I sit. But this problem belongs to every specialty, and solving it will require clinicians, scientists, engineers, health systems, regulators, and industry working together.

Next: why “it worked in the paper” might be the least useful sentence in clinical AI, and what we should be asking instead.

This post was prompted in part by Arvind Narayanan’s talk and essay, “What Will Be Left For Us to Work On?” His framing of how AI shifts human work from execution toward evaluation and accountability is well worth reading..

About the Author

Dr. Rajendra Singh is a Professor of Pathology at the University of Pennsylvania and co-founder of PathPresenter. He serves as a member of the Digital and Computational Pathology Committee of the CAP, Editorial Board of the WHO for Classification of tumors, 5th Edition and the Board of Digital Pathology Association.

The Future of Pathology Part 3: The Delivery Problem and the Infrastructure of Intelligence

After Part 1 and Part 2, many of you asked how the argument ends. This is where it lands.

When a field reaches consensus, it usually means the real work is about to begin.

Over the past two essays, I argued that the greatest challenge in precision oncology is no longer discovery. We continue to identify powerful biomarkers, develop increasingly sophisticated AI models, and generate molecular insights at an unprecedented pace. Scientific discovery creates possibility. Delivery creates outcomes.

This week, the Digital Pathology Association released a comprehensive recommendation statement on the validation, implementation, and clinical application of artificial intelligence in digital pathology. What struck me was not what was new, but what it quietly confirmed. The document is nominally about artificial intelligence. What it really describes is the operating environment required for precision medicine.

Three recommendations reveal just how much the conversation has changed.

1. Variability

The first concerns scanner variability. The article is explicit: algorithm performance depends on the scanner on which it is deployed, and validation must be demonstrated for each platform. This is more than a technical consideration and an acknowledgment of operational reality. Hospitals rarely operate identical hardware, workflows, or laboratory environments, and precision medicine must succeed despite that variability, not because it has been eliminated.

2. Oversight

The second concerns pathologist oversight. The recommendation is unequivocal: AI augments clinical expertise but does not replace it. In daily practice, this feels less like a regulatory requirement than a recognition of clinical reality. While AI excels at narrowly defined tasks, pathologists must integrate morphology, clinical history, molecular findings, rare disease patterns, and diagnostic judgment that no training set fully anticipates. The future is not autonomous diagnosis. It is intelligently augmented expertise.

3. Clinical Utility

The third concerns clinical utility, and this is where the bar is raised most significantly. A high-performing model is not enough. Clinical utility requires evidence that using the result actually changes patient management and improves outcomes. In dermatopathology, where a melanoma grading decision can determine treatment intensity, that distinction is not abstract. It moves the conversation beyond developing algorithms to delivering measurable clinical value, and that is a fundamentally harder problem.

Taken together, these recommendations describe something much larger than AI governance. They describe the infrastructure required for intelligence to become healthcare.

We have seen this pattern before. GPS technology existed for decades, yet it did not transform everyday life simply because satellites became more accurate. It became indispensable only when maps, smartphones, cloud computing, wireless networks, and real-time traffic data converged into a seamless ecosystem. The breakthrough was not a better satellite. It was the infrastructure that allowed millions of people to benefit from the intelligence those satellites had been generating all along. Precision oncology is approaching a similar inflection point. The biomarkers exist, the algorithms exist, and our biological understanding continues to expand. What remains inconsistent is the operational layer that validates performance across diverse environments, orchestrates workflows, connects fragmented systems, governs quality, and delivers the right intelligence at the precise moment a clinical decision must be made.

The conversation is evolving. We are moving from asking “Can AI detect disease?” to asking “What system ensures that AI improves care, every case, every day?” That is a much more difficult question, and it is the one that will determine whether precision medicine scales beyond a handful of highly resourced institutions or becomes accessible to patients everywhere.

Discovery will continue to accelerate. But the breakthroughs that matter are the ones that become routine, and that requires infrastructure, execution, and a field willing to treat delivery as seriously as discovery.

About the Author

Dr. Rajendra Singh is a Professor of Pathology at the University of Pennsylvania and co-founder of PathPresenter. He serves as a member of the Digital and Computational Pathology Committee of the CAP, Editorial Board of the WHO for Classification of tumors, 5th Edition and the Board of Digital Pathology Association.

What ASCO Taught Me About the Future of Pathology – Part 2

TROP2, Tissue, and the Missing Spatial Dimension

Last week, I wrote that the future of pathology would not be defined by any single technology—not AI, not molecular testing, and not digital pathology alone. Rather, the future belongs to those who can read tissue morphology, imaging, molecular biology, and clinical outcomes as a single integrated story rather than four separate chapters.

The discussions that followed were illuminating. Conversations with radiologists, oncologists, data scientists, and colleagues across the pharmaceutical industry reinforced something I had already begun to suspect: many of the most important questions in precision medicine no longer sit comfortably within the traditional boundaries of a single specialty.

As I reflected on the clinical data presented at ASCO this year, one story stood out because it illustrates exactly where this convergence may be heading. That story is TROP2.

Why TROP2 Matters: The Architecture of an ADC

For those not immersed in oncology drug development, TROP2 (trophoblast cell-surface antigen 2) is a transmembrane glycoprotein expressed across a wide range of epithelial malignancies, including lung, breast, bladder, and endometrial carcinomas. It has become a major target in oncology because it serves as an address for a rapidly expanding class of therapies known as antibody-drug conjugates (ADCs).

The concept is elegant. An antibody binds a target on the cancer cell, carries a highly potent cytotoxic payload, and delivers that payload into the cell following binding and internalization. In practice, many ADCs—including TROP2-directed therapies—also exert a “bystander effect,” where payload can diffuse into neighboring cells, making therapeutic activity less dependent on perfectly uniform target expression.

At ASCO, clinical data across multiple programs continued to reinforce the potential of this strategy. Early-phase and emerging late-phase results suggest meaningful clinical activity for TROP2-directed ADCs, including in combination with immunotherapy in selected settings. With multiple global trials underway across tumor types, this represents one of the most active areas of drug development in oncology today.

There is clear biology here—and real benefit for patients who respond. Yet as these data accumulate, a fundamental question becomes harder to ignore: how do we determine which patients will benefit most? The answer may expose a broader limitation in how we currently define biomarkers.

What the H-Score Misses

Today, our approach to TROP2 remains centered on expression. Pathologists stain tissue, evaluate intensity and extent, and report a score. Computational approaches have improved on this by quantifying expression more reproducibly and, in some cases, distinguishing subcellular localization.

These are meaningful advances. But they are still asking the same core question: What is the biology of TROP2?

At the microscope, the question often feels different: What is the biology surrounding TROP2?

A biomarker is never seen in isolation. Tumor cells exist within a structured environment: immune cells attempting to infiltrate, fibroblasts shaping stromal architecture, and physical barriers organizing the tumor into compartments. The tumor–stroma interface—the boundary between malignant epithelium and host response—is often where critical biology resides.

Experienced pathologists recognize these patterns intuitively. When I am teaching residents, I remind them that the tissue is often smarter than the pathologist. If the stroma is walling something off, that reaction is telling you something. The interface between tumor and host is rarely silent. Yet much of this contextual information is lost when complex tissue architecture is reduced to a single expression score.

That raises an important possibility: What if response to TROP2-directed therapy depends not only on target expression, but also on the biological context in which that target exists?

The Missing Spatial Dimension: An Ecosystem Approach

Consider the immune microenvironment.

A tumor with high TROP2 expression but minimal immune infiltration may behave very differently from one with the same expression profile embedded within an organized immune response. The molecular target may be identical, but the surrounding ecosystem is not.

Preclinical and early clinical observations suggest that ADC activity may interact with the tumor microenvironment in ways that extend beyond direct cytotoxicity. If those interactions contribute to clinical response, then baseline tissue architecture—including immune organization—becomes an important variable that current biomarker strategies only partially capture.

This concept is not new. In colorectal cancer, the Immunoscore demonstrated that the spatial distribution of immune cells—particularly at the invasive margin—has strong prognostic value and has influenced how we think about tumor–immune interactions more broadly.

Yet we have only begun to ask whether similar spatial features influence response to ADCs such as those targeting TROP2.

Even at a purely physical level, tissue architecture may matter. A TROP2-high tumor with dense stromal compartmentalization could experience very different drug delivery and payload diffusion dynamics than a tumor with a more permissive structure. These are measurable features—but they are rarely incorporated into current biomarker frameworks.

This points toward a broader realization: we may have spent years searching for answers in individual molecules when the more informative signal, in some settings, lies in the relationships between them—their spatial arrangement, biological context, and interaction within a tissue ecosystem.

Beyond the Slide

The second observation that stayed with me after ASCO is that biology does not operate at a single scale.

Tumors do not live on glass slides. They exist within tissues, tissues within organs, and organs within patients. It follows that response to therapy is shaped across all of these levels simultaneously.

Some relevant features are visible histologically. Others emerge through imaging, molecular profiling, or longitudinal clinical data. Increasingly, computational approaches are making it possible to integrate these layers.

The central question for precision oncology is no longer which modality is most informative in isolation. It is whether these modalities are capturing different dimensions of the same biological system—and whether we are building the infrastructure to interpret them together.

For decades, medicine has organized these data into separate workflows, databases, and specialties. Biology never made those distinctions.

The Bigger Lesson

TROP2 is not the story. It is a case study.

For the past two decades, precision medicine has focused on identifying the right molecule—and that approach has transformed cancer care. But in some contexts, the molecule alone may not be sufficient to explain therapeutic response.

Tumors exist within tissues. Tissues exist within organs. Organs exist within patients. Every level contains information. Every level influences outcome.

This is why pathology is becoming more important—not less—in the era of AI and advanced molecular diagnostics. Not because it replaces other disciplines, but because it is inherently grounded in biological context.

The role of the pathologist is evolving accordingly. It is no longer limited to interpreting a slide or assigning an expression score. It increasingly involves connecting tissue architecture, molecular data, imaging, and clinical outcomes into a more complete model of disease.

That integration is beginning to take shape. The groups that build this connective infrastructure first will be positioned to answer some of the most important biomarker questions in oncology.

The tissue has been telling us this story for years. Imaging has been telling the same story from a different vantage point. Molecular data and clinical outcomes add additional layers. For most of modern medicine, we have read these stories separately.

We are finally learning how to read them together.

About the Author

Dr. Rajendra Singh is a Professor of Pathology at the University of Pennsylvania and co-founder of PathPresenter. He serves as a member of the Digital and Computational Pathology Committee of the CAP, Editorial Board of the WHO for Classification of tumors, 5th Edition and the Board of Digital Pathology Association.

What ASCO Taught Me About the Future of Pathology

A field report from the world’s largest oncology meeting—and a call to my fellow pathologists.

This weekend, I joined tens of thousands of oncologists, researchers, pharmaceutical executives, and data scientists at the ASCO Annual Meeting. The entire cancer medicine ecosystem was gathered in one place.

Yet among the thousands of people discussing the future of cancer care, I could count the pathologists on two hands.

I sat through presentations on antibody-drug conjugates (ADCs), immunotherapy combinations, spatial biomarkers, and novel companion diagnostics. Different diseases. Different companies. Different therapeutic strategies. Yet I found myself writing the same note repeatedly in the margin of my notebook:

This depends on pathology.

The more I listened, the more obvious it became that many of oncology’s biggest challenges are no longer drug problems. They are patient-selection problems. And patient selection begins with understanding tissue.

The industry is chasing the biology that pathologists have spent their careers studying.

For the first time in my career, I believe pathology is positioned not only to diagnose disease, but to help determine treatment.

1. The Industry Is Chasing Tissue Biology

No matter the disease or the drug class, the central question in oncology today is patient selection. Who will respond? Who will benefit? Who should receive this therapy?

Whether the discussion involved checkpoint inhibitors, ADCs, or cellular therapies, the answers do not live in a liquid biopsy or a genomic sequence alone. They live in tissue architecture inside the tumor microenvironment, within the spatial relationships between malignant cells and immune infiltrates.

The answers oncologists are looking for reside in tissue. And that means pathology must be at the table.

2. The H&E Slide Is Far Richer Than We Imagined

The glass slide sitting under our microscopes contains a vast reservoir of unextracted information. Spatial arrangements of immune cell subsets, subtle variations in stromal organization, cellular heterogeneity across a specimen are not incidental features. They are biological signals that influence treatment response and clinical outcomes.

Digital pathology, computational modeling, multiplex imaging, and AI allow us to quantify what was previously qualitative, transforming what we have always observed into measurable, reproducible biological data.

These technologies are not replacing the pathologist’s eye. They are extending its reach. This is not a disruption to pathology; it is pathology at greater resolution.

3. Pathology’s Second Act: From Diagnosis to Treatment

For most of my career, pathology has answered the first question in cancer care: What disease does this patient have? Oncology is now asking a second question: What treatment is most likely to help this patient?

That question increasingly depends on information embedded within the tissue itself:

  • Why does one patient respond to immunotherapy while another progresses?
  • Which features of the tumor microenvironment predict resistance?
  • What makes a targeted therapy effective in one patient and ineffective in another?

These are pathology questions. And they are among the most consequential questions in medicine today.

4. Why Pathologists Must Lead

No algorithm inherently understands tissue biology. No software comprehends the clinical weight of a diagnostic error. Pathologists do. That is precisely why we must be the ones evaluating, validating, and challenging these tools, separating what is biologically meaningful from what is merely statistically interesting.

The pathologist of the next decade will need an expanded toolkit. We must learn how to interpret spatial biology, understand AI-derived biomarkers, engage with computational pathology, and participate in the development of the next generation of companion diagnostics. This is not about abandoning what we know; it is about building on it.

The future of oncology cannot be built by software engineers or clinical oncologists alone. It requires pathologists who speak both the language of tissue and the language of data.

The future of pathology will not be determined by artificial intelligence. It will be determined by whether pathologists choose to engage with it.

The tools are changing. The mission is not.

Defining the Future

When I return to ASCO next year, I hope I need more than two hands to count the pathologists in attendance. Not because we are defending our relevance, but because we are helping shape the future of cancer care.

The future of precision oncology is being built now.

And pathology belongs at the center of it.

About the Author

Dr. Rajendra Singh is a Professor of Pathology at the University of Pennsylvania and co-founder of PathPresenter. He serves as a member of the Digital and Computational Pathology Committee of the CAP, Editorial Board of the WHO for Classification of tumors, 5th Edition and the Board of Digital Pathology Association.

A Biomarker by Any Other Name Would Smell as Sweet: Building an AI-Powered Internal Concept Library

Part two of a two-part series on the idea of Internal Concept Libraries. Read part one.

Building an AI-powered internal concept library

The promise of artificial intelligence in pathology is often framed around image analysis and diagnostic augmentation. But one of the most immediate opportunities may lie elsewhere: transforming decades of unstructured pathology reports into computable, searchable clinical data.

A recent pathology informatics project by Raj Singh, Professor of Pathology at UPenn and PathPresenter co-founder, and Alexander Goel, COO at PhenoML, explored this challenge through the development of an AI-driven extraction pipeline designed around a central principle: preserve all clinically meaningful observations, even when standardized terminology systems fail to represent them completely.

The project focused on constructing an Internal Concept Library capable of bridging the gap between narrative pathology reporting and structured research databases.

The dataset consisted of 1,155 free-text pathology reports drawn from The Cancer Genome Atlas (TCGA), including breast, lung, and prostate cancer cases spanning 1978 to 2013. These reports largely predated modern synoptic reporting practices and therefore represented the type of highly variable narrative text still common throughout historical pathology archives.

The objective was not simply to extract data fields. It was to evaluate whether a concept-first architecture could preserve more clinically useful information than traditional terminology-dependent pipelines.

The system architecture was intentionally designed to separate clinical meaning from terminology standardization.

The pipeline began with ingestion of raw pathology reports. Cases were first routed through disease-specific classification using external terminology APIs and machine learning classifiers. Each report was then processed using large language model prompts tailored to individual disease sites.

Rather than allowing unconstrained extraction, the prompts used a closed-world design philosophy. The model was explicitly instructed that the field list was exhaustive, preventing the AI from hallucinating expected findings that were not actually documented in the report.

Each extracted observation was assigned an internal concept identifier before any attempt at external terminology mapping occurred. This architectural decision represented the core innovation of the project.

In traditional pipelines, extraction is often tightly coupled to successful SNOMED encoding. If no suitable standard code is found, the observation may be excluded from downstream datasets entirely.

In the concept-library model, however, the internal concept persists regardless of SNOMED availability.

Only after extraction were findings submitted to terminology matching systems to identify corresponding SNOMED CT or OMOP standard concepts. When no standard code existed, custom OMOP concept IDs were assigned instead, preserving full queryability within the OMOP ecosystem. The results highlighted the scale of the terminology gap.

Study results
Results of extraction and mapping

Across 1,155 reports, the system extracted 13,323 clinical observations. Of these, approximately 59.6% successfully mapped to SNOMED CT concepts. Roughly 39.1% could not be mapped to standard terminology despite representing clinically meaningful findings. An additional 1.3% were flagged as uncertain due to ambiguity or conflicting evidence within the report text.

Under conventional code-first architectures, a substantial portion of these observations would likely have been excluded from structured datasets entirely. Instead, the Internal Concept Library preserved all extracted findings.

The project also evaluated extraction completeness against the College of American Pathologists electronic Cancer Protocols (CAP eCPs), using the CAP templates as a formal benchmark for information coverage. This approach itself represented a novel methodological contribution, as CAP templates have rarely been used as explicit completeness standards in published large language model pathology extraction research.

Coverage varied by disease site. Breast reports achieved approximately 32.9% CAP field extraction, prostate reports 36.5%, and lung reports 21.8%, with an overall completeness rate of 29.2%.

At first glance, those percentages may appear modest. But the context matters enormously.

The TCGA reports originated from an era before structured synoptic reporting became widespread. Many CAP-defined fields simply did not exist within the source narratives. In other words, lower completeness often reflected absence of documentation rather than extraction failure.

The researchers hypothesized that applying the same pipeline to contemporary synoptic or semi-structured pathology reports would likely produce substantially higher completeness rates.

Several important design lessons emerged from the project.

  • First, closed-world prompting proved essential. Explicitly constraining the extraction schema reduced over-extraction and prevented the model from inferring clinically plausible but undocumented findings.
  • Second, the team found that self-reported LLM confidence scores were poorly calibrated and unreliable for quality assurance. Instead, better performance indicators emerged from structured assertion logic (e.g. forcing every extracted finding into explicit states such as present, absent, or uncertain rather than allowing vague narrative interpretation), explicit review of terminology concordance (verifying that extracted findings align consistently with accepted clinical vocabularies), and cross-field consistency checks (evaluating whether extracted observations remain clinically coherent when considered together).
  • Third, the researchers intentionally avoided creating separate post-extraction normalization stages. Additional normalization layers often compound error propagation while duplicating functionality already present in terminology APIs.

Most importantly, the study demonstrated that pathology AI systems should prioritize preservation of meaning over immediate standardization.

This distinction has major implications for clinical trial recruitment and translational research. Using concept-level querying, researchers could identify highly specific patient cohorts using combinations of biomarkers, grading systems, and morphologic findings regardless of whether standardized terminology mappings existed for every observation.

Queries such as:

  • Triple-negative breast cancers with Ki-67 greater than 50%
  • Gleason score ≥8 with perineural invasion

could operate across the full extracted dataset, not merely the subset successfully encoded in SNOMED.

The project also pointed toward future directions in ontology management itself. Singh and Goel are considering multi-agent AI frameworks capable of automatically reviewing new observations, searching multiple ontologies, drafting candidate mappings, and escalating only genuinely ambiguous cases for human expert review. In such systems, pathologists and informaticians would spend less time performing repetitive terminology lookups and more time adjudicating clinically meaningful ambiguity.

Internal Concept Library multi agent framework

Ultimately, the work reinforces a broader shift occurring across healthcare AI. The goal is no longer simply extracting structured data from pathology reports. The larger challenge is preserving clinical meaning at scale while enabling interoperability, research, and machine reasoning.

Pathologists have always documented nuanced clinical interpretation for human readers. Building systems that preserve that same nuance for computational systems may become one of the defining informatics challenges of precision medicine.

What’s in a Namespace: The Critical Role of an Internal Concept Library for Scalable Pathology Data Management

Part one of a two-part series on the idea of Internal Concept Libraries

Internal Concept Library

Modern pathology departments generate enormous volumes of clinically rich information every day. Yet much of that information remains effectively invisible to the systems used for identifying patients for research studies, biomarker-driven therapies, and clinical trials.

Much of the problem stems from a lack of structure.

Across healthcare institutions, pathology reports are still dominated by unstructured narrative text: highly expressive for human interpretation, but difficult for computers to interpret consistently. As precision medicine increasingly depends on identifying patients with highly specific molecular and histologic characteristics, this gap between human-readable and machine-readable pathology is becoming one of the most important bottlenecks in translational research. It prevents modern tools from leveraging critical data at scale.

At the 2026 Association for Pathology Informatics Summit, Raj Singh, Professor of Pathology at UPenn and PathPresenter co-founder, asked the audience to consider a clinical trial scenario: A researcher needs to identify patients with triple-negative breast cancer whose Ki-67 proliferation index exceeds 20%. Querying their institution’s database using the standard SNOMED CT concept associated with Ki-67 expression returns 47 patients.

The problem? The actual number of eligible patients in the database is 112.

Sixty-five patients were missed—not because the information was absent, but because it was described differently across pathology reports.

One report may say “Ki-67: 35%.” Another may describe “Proliferation index: 45%.” Others may reference “MIB-1 labeling index,” “Ki67 high,” or “Brisk mitotic activity.” To a pathologist, these phrases convey closely related clinical meaning. To a conventional search system relying on standardized terminology mapping, many are effectively invisible.

Ki-67 search results

This challenge is not unique to Ki-67. It reflects a broader reality in pathology reporting.

Different institutions, and even different pathologists within the same institution, use varying terminology, abbreviations, units of measurement, and narrative styles. A tumor described as “2.2 cm” in one report may appear as “22 mm” in another and “0.022 m” in a third. Humans easily recognize these as equivalent observations. Most database search tools do not.

The industry’s primary solution to this problem has been terminology standardization frameworks such as SNOMED CT and the OMOP Common Data Model (CDM). These systems are essential foundations for interoperability and large-scale clinical research, and they are both impressive and commendable. But they also expose an important limitation: medicine evolves faster than standardized terminology systems can fully capture.

SNOMED CT (Systematized Nomenclature of Medicine Clinical Terms) is one of the world’s most comprehensive clinical vocabularies. It provides unique identifiers for clinical concepts so that systems across institutions can interpret observations consistently. When a pathology finding is successfully mapped to SNOMED, that information becomes interoperable and searchable across research networks.

OMOP CDM, meanwhile, provides a standardized database structure widely used by the OHDSI research community. OMOP enables federated research across hundreds of institutions while preserving patient privacy. Importantly, OMOP also supports custom local concepts through reserved concept ID namespaces. In other words, it explicitly leaves space in its system for concepts that haven’t been defined yet, knowing that medicine and research are always expanding. 

Together, SNOMED and OMOP form the backbone of many modern clinical data infrastructures. But they are not sufficient on their own. 

One core issue is that traditional pathology extraction pipelines often treat successful terminology mapping as a prerequisite for data preservation. If an extracted observation cannot be confidently mapped to a standard code, it is frequently discarded from downstream datasets entirely. Even if the data is preserved in local OMOP custom concepts, those local concepts do not provide any shared meaning across institutions.

This creates a dangerous form of silent data loss. The pathologist documented the finding correctly. AI extraction systems may even identify it correctly. But if no standard code exists, or if terminology matching fails, the information disappears from searchable research datasets.

For institutions engaged in clinical trials, biomarker research, or retrospective cohort studies, this means potentially eligible patients are never identified.

The solution proposed by Dr. Singh, in collaboration with Alexander Goel, COO at PhenoML, is the development of an Internal Concept Library. An internal concept is one stable name for one clinical observation: it gives a consistent way for people and systems to recognize it, regardless of how the pathologist wrote it. An Internal Concept Library acts as a canonical semantic layer between unstructured pathology language and external terminology systems.

For example, multiple report expressions such as:

  • “ER positive”
  • “ER 3+”
  • “Strongly reactive estrogen receptor”
  • “Allred score 8”

can all resolve to a single internal concept:

er_status = positive

The internal concept becomes the persistent primary key. Standard terminology mappings are then attached downstream when available. Instead of treating SNOMED codes as the primary identifiers, this approach first maps observations to stable internal concepts designed specifically around clinical meaning.

This architectural distinction is critically important. If SNOMED mapping succeeds, the observation gains full interoperability. If SNOMED mapping fails, the observation is still preserved, queryable, and available for cohort discovery. Nothing is lost.

This approach also enables institutions to absorb local terminology variation across multiple sites and organizations to enable federated queries without constant schema redesign. New biomarkers, evolving grading systems, and emerging molecular concepts can be added incrementally to the concept library without restructuring the entire database architecture. New sites can be added to the collective queryable data network.

For pathology departments and institutional researchers, the implications are substantial.

An Internal Concept Library can improve cohort identification accuracy, preserve historically inaccessible clinical detail, support retrospective enrichment of older pathology archives, and create a more resilient foundation for AI-assisted research workflows.

Equally important, it aligns with how pathologists actually communicate clinically meaningful information. Pathologists have always preserved nuance for human readers. The challenge now is preserving that nuance for machines.

As precision oncology continues to advance, institutions that rely solely on rigid terminology (or no consistency at all) matching risk overlooking large portions of their own clinical knowledge base. The future of scalable pathology informatics may depend less on forcing every observation into a predefined standard, and more on building systems capable of preserving meaning first, then standardizing second.

The shift from code-first architecture to concept-first architecture may prove essential for the next generation of clinical research and patient discovery.

Part Two: Building a Library with AI

In the second part of this series, we’ll look at the design and results of a real-world project to build an internal concept library using a customized AI pipeline. Read Part Two: A Biomarker by Any Other Name Would Smell as Sweet: Building an AI-Powered Internal Concept Library

Why Are We Still Shipping Glass Slides in 2026?

by Patrick Myles
CEOr, PathPresenter

Every day, patients rely on pathology second opinions to confirm diagnoses, guide treatment decisions, and connect them with subspecialty experts who may be located across the country, or across the world.

And yet in 2026, many of these consultations still begin the same way they did decades ago: a cardboard slide box. Bubble wrap. FedEx.

The expertise has never been the problem. The logistics are.

Why Second Opinions Matter

Pathology is increasingly specialized. A community pathologist may only encounter a rare tumor type a few times in their career, while subspecialty experts at academic centers review those cases every day.

Second opinions improve patient care by connecting physicians and patients with the right expertise, regardless of geography. But the workflow supporting these consultations remains heavily dependent on physical glass slides moving through shipping networks.

A typical external consult can still take 5–14 days—not because the diagnosis is difficult, but because the slides are traveling.

Along the way, there are familiar pain points:

  • Shipping delays or lost slides
  • Manual transfer of patient metadata
  • Disconnected systems
  • Multiple handoffs between institutions
  • Additional delays for stains or molecular testing

Even when slides are digitized, many organizations are still relying on generic file-sharing tools with no integrated workflow, audit trail, or LIS connectivity. In many cases, we simply replaced FedEx with Dropbox.

Digital Pathology Changes the Equation

Digital pathology allows whole slide images to be securely shared instantly with experts anywhere in the world.

That means:

  • Faster turnaround times
  • Improved collaboration
  • Better patient access to subspecialty expertise
  • Reduced operational friction

At PathPresenter, we recognized that remote consultation workflows require more than image viewing alone. They require infrastructure. That’s why we built a vendor-agnostic platform designed specifically for digital pathology consultations:

  • Integrated viewing of pathology, radiology, and clinical images.
  • Secure PHI-enabled collaboration
  • Structured consult workflows
  • Full audit trails
  • LIS integration
  • Async collaboration between institutions
  • Integrated viewing of pathology, radiology, and clinical images
Consult web portal diagram

PathPresenter has a secure, scalable, and HIPAA compliant digital workflow for clinical consults and referral

Most importantly, we designed the system to work across scanners, storage platforms, AI applications, and laboratory system because pathology is never a single-vendor environment.

Making Digitization Standard Operating Procedure

One of the biggest opportunities ahead is extending digital pathology beyond large academic centers and into outreach and referral hospitals.

A fully digital consultation workflow only works if cases can originate digitally. Historically, barriers like scanner costs, storage requirements, deployment complexity, and interoperability challenges slowed adoption. But cloud-native platforms, zero-footprint deployment models, and improved interoperability are rapidly changing the equation.

The transition to digital pathology will require collaboration across the industry. The good news is that this collaboration is finally beginning to happen, with many highly capable but lower cost scanners available.

The Future of Pathology Shouldn’t Depend on FedEx

Patients should not wait days or weeks for access to subspecialty expertise because slides are sitting in transit. The technology already exists to make pathology consultations digital, connected, and immediate.

Now it’s time for the industry to work together, and with partners at institutions and outreach hospitals to scale it.

Looking to Digitize your Consultation Workflow?

If your institution is thinking about expanding digital pathology consultations, enabling remote second opinions, or building a more connected pathology network, we’d love to talk. Reach out to the PathPresenter team to learn how we’re helping institutions modernize pathology workflows and expand access to subspecialty expertise worldwide.


About the Author

Patrick Myles CEO of PathPresenter. Previously, he was CEO of Huron Digital Pathology, and Vice President of Business Development for Teledyne DALSA. He served as a board member of the Digital Pathology Association.

The algorithm is ready. The question is whether our departments are.

As a pathologist, this feels to me like one of the most important moments our specialty has seen in decades.

Not because AI in pathology is new. Most academic departments have been thinking about computational pathology, biomarker quantification, and digital workflows for the last few years. What feels different now is that these tools are no longer being discussed primarily as research infrastructure or future possibilities. They are beginning to shape real treatment decisions.

Roche’s acquisition of PathAI is a significant signal that computational pathology has moved into the center of precision oncology strategy. The implications are larger than a single acquisition.

For years, pathologists have understood something that the broader healthcare ecosystem is only now fully appreciating: the biopsy is not simply supporting information for oncology care. Increasingly, it is where therapeutic eligibility is determined. As AI-enabled companion diagnostics become regulatory-grade clinical tools, pathology workflows are becoming directly connected to treatment access.

The emerging generation of ADCs and biomarker-driven therapies illustrates this clearly. Many of these treatments depend on increasingly nuanced tissue-based interpretation — quantitative scoring, spatial context, low-expression biomarkers, and patterns difficult to assess reproducibly through conventional microscopy alone. The pathologist remains central to interpretation, but the infrastructure around that interpretation is changing rapidly.

The Real Challenge: Deployment Over Development

An algorithm may be developed in a research environment, but it ultimately it has to function inside a clinical department. It must integrate into the daily workflow of pathologists, connect to existing laboratory systems, support regulatory and reporting requirements, and operate within institutions that have already made long-term infrastructure decisions. That deployment challenge is likely to become one of the defining questions of computational pathology over the next several years.

The academic medical centers where many of these patients are diagnosed are extraordinarily complex environments. They are simultaneously delivering clinical care, running trials, training residents and fellows, conducting translational research, supporting tumor boards, and managing large-scale image archives. Any AI companion diagnostic that hopes to achieve broad adoption will need to fit naturally into that ecosystem.

This is why workflow and institutional trust matter as much as algorithmic performance.

Pathology departments do not adopt infrastructure lightly. The systems that succeed over time are the ones that integrate cleanly into clinical operations, support a wide range of departmental needs, and earn confidence through years of daily use. In practice, the operational realities of pathology often determine whether even highly sophisticated technologies can meaningfully reach patients.

None of this diminishes what companies like Roche and PathAI have accomplished. The field needs validated, clinically deployable AI systems, and the progress being made is genuinely exciting for pathology and oncology alike.

But the next phase of the field may be defined less by whether these algorithms can be built, and more by how effectively they can be deployed within the institutions where precision oncology is actually practiced.

About the Author

Dr. Rajendra Singh is a Professor of Pathology at the University of Pennsylvania and co-founder of PathPresenter. He serves as a member of the Digital and Computational Pathology Committee of the CAP, Editorial Board of the WHO for Classification of tumors, 5th Edition and the Board of Digital Pathology Association.

Success Story: Expanding Access to Expert Cancer Diagnostics Through Digital Pathology Consultations

How The Ohio State University Comprehensive Cancer Center and PathPresenter enabled faster, more equitable second opinions across a regional network

PathPresenter IMS at OSUCCC

In Brief

Access to subspecialty pathology expertise can significantly influence cancer diagnosis, treatment planning, and patient outcomes. Yet many regional and community hospitals lack the depth of specialized pathology resources available at large academic medical centers. The Ohio State University Comprehensive Cancer Center-Arthur G. James Cancer Hospital and Richard J. Solove Research Institute (OSUCCC-James) addressed this gap by launching a digital pathology second opinion consult service across its James Cancer Network.

By partnering with PathPresenter, OSUCCC-James implemented a scalable, digital-first consultation workflow that eliminates traditional barriers such as physical slide transport, redundant data entry, and incompatible systems between institutions. At the center of this solution is PathPresenter’s clinical consultation module, enhanced with OCR-driven case accessioning that automatically extracts patient and case metadata directly from slide labels.

This innovation allows partner sites like Wooster Community Hospital to seamlessly submit cases for expert review without requiring a shared laboratory information system (LIS) or additional administrative burden. Subspecialist pathologists at OSUCCC-James can review cases within hours, dramatically reducing turnaround time compared to legacy workflows that could take days or weeks.

The result is a more connected, efficient, and equitable diagnostic network: one that extends advanced cancer expertise into local communities, improves clinical decision-making, and opens new pathways to clinical trials and specialized care.

The Goal: Bringing Specialist Expertise Closer to Patients

Accurate pathology interpretation is foundational to effective cancer care. Subtle diagnostic nuances can determine treatment pathways, eligibility for targeted therapies, and access to clinical trials. While academic centers like OSUCCC-James house world-class subspecialty expertise, many smaller or regional hospitals operate with limited pathology resources.

Recognizing this disparity, OSUCCC-James included remote pathology consultations in the services it offers through The James Cancer Network, a collaborative initiative designed to extend advanced oncology services beyond its main campus. Through this network, partner institutions gain access to training, research opportunities, and, critically, expert diagnostic support.

The need for consultations was clear. However, delivering this capability at scale required overcoming several operational and technical challenges. Among OSUCCC-James’s requirements for a successful solution:

  • Rapid turnaround times to support timely clinical decisions
  • Minimal logistical complexity, especially eliminating physical slide transport
  • Reduction of manual data entry and administrative overhead
  • Flexibility to integrate with diverse systems across partner institutions
  • A seamless and reliable experience for consulting pathologists
  • Strong privacy, security, and regulatory compliance
  • Scalability to support a growing network

Digital pathology was the obvious foundation, but implementing it effectively required more than simply digitizing slides.

The Challenge: Fragmented Systems and Inefficient Workflows

While digital pathology enables remote viewing of whole slide images (WSIs), the surrounding workflow, particularly case accessioning and data management, often remains fragmented.

In typical consultation workflows, when a regional hospital submits a case for consultation, there is plenty of work besides the actual slide scanning. Staff must create or export a case record from their local LIS, and often re-enter patient and case information into the receiving institution’s system.

This process introduces delays, increases the risk of transcription errors, and places additional strain on already limited staff resources. The problem is compounded when partner institutions use different LIS platforms or lack such systems altogether.

For OSUCCC-James, these inefficiencies posed a significant barrier to scaling their consultation program. They needed a solution that could:

  • Work across heterogeneous or absent LIS environments
  • Automate case creation without duplicating effort
  • Maintain data integrity and compliance
  • Fit naturally into existing clinical workflows

The Solution: PathPresenter’s Consultation Module with Automated Case Accessioning

To address these challenges, OSUCCC-James partnered with PathPresenter, leveraging a long-standing relationship built on educational, research, and clinical deployments of PathPresenter’s IMS.

PathPresenter’s platform already offered a robust, FDA-cleared clinical viewer and class-leading consultation capabilities. For this initiative, the collaboration focused on extending these capabilities to streamline remote case submission and management.

The key innovation was an OCR-based case accessioning workflow.

In this approach, PathPresenter extracts relevant metadata, such as patient identifiers and case details, directly from the labels present on the scanned slide images. This information is automatically interpreted and structured into a digital case record, eliminating the need for manual data entry.

This automated workflow has several immediate advantages:

  • No reliance on a shared LIS or complex integrations: Partner institutions do not need to use the same system as OSUCCC-James
  • Reduced administrative burden: Staff are not required to manually enter patient and case details
  • Improved accuracy: Automated data capture reduces transcription errors
  • Faster case submission: Cases can be created and transmitted in minutes

Once a case is submitted, the digital case is automatically ingested into OSUCCC-James’ central EPIC LIS and assigned to the requested subspecialist pathologist or an appropriate subspecialist based on case requirements. The pathologist reviews the case using PathPresenter’s IMS and FDA-cleared viewer, and signs out the diagnosis, with the report automatically transmitted back to the LIS and the referring hospital, closing the loop.

This end-to-end workflow transforms what was once a multi-day or multi-week process into one that can be completed within hours.

With this program’s implementation, announced by OSUCCC in April, remote centers like Wooster Community Hospital can now submit cases without needing an LIS at all, relying instead on the automated workflow built into the platform to handle the flow without increasing the administrative burden.

Results and Impact: Faster Diagnoses, Broader Access, Better Care

The implementation of PathPresenter’s digital consultation workflow brings measurable impact across the James Cancer Network.

Dramatically Reduced Turnaround Times
Consultation timelines can reduce from days or weeks to just hours. This acceleration enables faster clinical decision-making, which is especially critical in oncology where timely intervention can significantly affect outcomes.

Elimination of Physical Logistics
The need to transport glass slides has been effectively removed. This not only reduces delays but also minimizes the risk of damage or loss during transit.

Streamlined Workflows and Reduced Administrative Burden
Automated case accessioning eliminates redundant data entry, freeing up staff to focus on higher-value clinical tasks. It also improves consistency and data quality.

Interoperability Across Diverse Systems
The solution accommodates a wide range of partner institutions, regardless of their existing IT infrastructure. This flexibility is essential for scaling across a regional network.

Enhanced Pathologist Experience
Consulting pathologists can review cases within a familiar, integrated environment. The workflow aligns with existing diagnostic practices, ensuring efficiency without disruption.

Strong Security and Compliance
The platform meets stringent requirements for privacy protection, regulatory compliance, and data security, critical factors in handling sensitive patient information.

Scalable Foundation for Network Growth
With a proven digital infrastructure in place, OSUCCC-James can continue expanding its network, bringing advanced diagnostic capabilities to more communities.

Enabling the Future of Connected Cancer Care

Beyond operational improvements, the broader impact of this initiative lies in its ability to connect patients to more advanced care pathways.

By making subspecialty expertise readily accessible, the program ensures that patients in regional hospitals receive the same level of diagnostic insight as those treated at a major academic center. This can influence not only diagnosis but also eligibility for targeted therapies and clinical trials.

Importantly, it also strengthens collaboration between institutions. Community hospitals become more deeply integrated into the academic center’s ecosystem, benefiting from shared knowledge, resources, and innovation.

As OSUCCC-James continues to expand the James Cancer Network, digital pathology consultations serve as a cornerstone of its strategy, demonstrating how technology can bridge gaps in access and elevate the standard of care across an entire region.

Conclusion

The partnership between OSUCCC-James and PathPresenter illustrates how thoughtful application of digital pathology can overcome longstanding barriers in healthcare delivery. By focusing on automation, interoperability, and user-centered design, the solution transforms complex workflows into seamless, scalable processes.

The result is not just faster consultations, but a more equitable healthcare system, one where geography no longer limits access to expert cancer diagnostics.

More information

Interested in learning more about how solutions like this can help solve your own challenges? Contact us at pathpresenter.com/contact

Contact us

The Pathologist as the Architect of Clinical Intelligence

Last week, I was invited to serve as a judge for a “Shark Tank” style competition at the University of Pennsylvania. At the conclusion of several spectacular presentations, all of which featured the innovative use of AI, a resident in the audience asked the million-dollar question: “Do I have a future as a pathologist?”

It is a question often framed with uncertainty—will AI replace us or commoditize our work? My answer is that the role of the pathologist is not diminishing but quietly expanding. We are moving from a diagnostic service to becoming the most valuable data engine in medicine. For decades, pathology has been the “ground truth” where cells and microenvironments are observed directly. Yet, this richness has historically been a computational dead end because we digitized images without digitizing meaning.

That transition is occurring now through three converging technological shifts that turn the image management system into an active collaborator. First, Foundation Models now read whole-slide images directly to extract critical features like TIL density and tumor burden. Second, Structured Narrative tools utilizing large language models convert our narrative prose into governed data with accuracy exceeding manual abstraction. Finally, Agentic Frameworks allow for cohort definition in plain language, delivering traceable results in minutes and removing traditional operational bottlenecks. Together, these technologies ensure that everything visible on a slide becomes computable.

The true power of these technologies lies in their invisibility; intelligence is embedded directly into the Image Management System (IMS) rather than being siloed in a dashboard. While the pathologist focuses on the diagnosis, background agents handle the operational heavy lifting. Specimen Logistics agents follow specimens through the lab to identify delays or staining inconsistencies before they breach turnaround times. Predictive Metadata Enrichment anticipates clinical pathways by identifying likely immunostains or molecular tests, suggesting CPT codes, and pre-filling reports. By surfacing relevant literature and potentially providing “virtual stains” that eliminate 48-hour waits, the platform ensures expertise is supported by a comprehensive data layer without interrupting workflow. Even the complex ontologies and data standards that once took years of committee work to codify can now be harmonized on the fly. By utilizing supervisory agents that map local findings to global standards in the background, we ensure that a diagnosis rendered in one institution carries the same precise, computable meaning in another. This background alignment allows the pathologist to focus entirely on the tissue while the system ensures the resulting data is instantly interoperable and research-ready.

This evolution reframes our impact from “throughput” to “leverage”. Each diagnosis is no longer a discrete event; it becomes a computable data point in a multimodal dataset. Because we see the disease itself in its structural context, we are uniquely positioned to lead this transformation. The future of healthcare will be built around the data that pathologists already hold. Our opportunity is to remain operationally central by shaping how these systems are designed and how meaning is preserved. Pathology is no longer just a service; it is the most valuable data asset in medicine, and those who build this capability first will hold the keys to the future of personalized care.

The future of pathology has never been brighter; we are moving from the basement of the hospital to the very epicenter of precision medicine, where our insights will drive every major clinical decision in the digital age.

About the Author

Dr. Rajendra Singh is a Professor of Pathology at the University of Pennsylvania and co-founder of PathPresenter. He serves as a member of the Digital and Computational Pathology Committee of the CAP, Editorial Board of the WHO for Classification of tumors, 5th Edition and the Board of Digital Pathology Association.