Deep learning can help pathologists with their core task: interpreting microscopic images. Over the past decade, computational pathology has advanced rapidly, but the traditional product-development playbook, training a new model largely from scratch for each use case, made progress slower and more expensive than it needed to be.
Pathology is not one problem; it is thousands. The work spans the cellular (detecting mitotic figures, infectious organisms and cell types), the tissue level (finding foci of invasion), and the case level (integrating findings across multiple slides and specimens). Beyond describing what is visible, many high-value questions are predictive: In some settings, what molecular alterations might be present? How aggressive is the disease? Will this patient respond to a therapy?
Historically, each new AI tool required its own curated training and validation cohorts and large volumes of pathologist annotations: often hundreds of thousands of labeled regions. That approach does not scale to the breadth of tasks across anatomic pathology. Foundation models change the economics and the pace of development.
A useful way to think about foundation models is as a core layer in a larger product stack, not the product itself. A whole slide image (WSI) is first converted from raw pixel space into information-rich embeddings by a foundation model. Those embeddings become a shared input to many smaller task-specific models, each trained to answer a particular clinical question – like detecting tumor, grading dysplasia, or quantifying a biomarker.
The AI product is what orchestrates this end-to-end: it tiles the WSI, generates embeddings, applies one or more task models, summarizes results at the slide or case level, and presents them through a workflow-integrated interface that pathologists can trust and use. Foundation models can accelerate development and improve performance across tasks, but clinical value comes from the full product: robust pipelines, validation, integration, and user experience.
Improving the development process
Foundation models are trained on large, diverse collections of unlabeled WSIs using self-supervised learning. Rather than relying on exhaustive expert labels upfront, the model learns useful visual representations by solving prediction or contrastive tasks on the data itself. That reduces the cost and friction of building a strong starting point.
After pretraining, the foundation model converts each image patch into an embedding, which is a compact numerical representation that captures salient histologic patterns. Downstream models can then operate on embeddings rather than raw pixels, which typically reduces the amount of labeled data needed to reach high performance.
What are embeddings?
Embeddings are how computers translate human-centric data (like images or text) into a structured numerical format they can work with.
One way to picture this is a gallery organized by similarity: images of dogs cluster together; puppies sit nearby; cats are close, but distinct.
When you ask for “fluffy pets,” you search the neighborhood where those concepts live rather than matching exact words.
Embeddings are the coordinates that make that organization possible.
In practice, this means you are not starting each project from scratch. You start from rich embeddings and fine-tune smaller task-specific models for detection, quantification, grading, or prediction. The result is faster iteration, lower annotation burden, and a development process that can support a broad portfolio of tools.
Building better models for pathology
Foundation models make it feasible to build AI tools across the workflow – tools that can add value on a large fraction of slides, not only in narrow niches. In oncology, AI tools can support screening and triage, diagnostic decision support, grading and staging, quantitative biomarker readouts for companion diagnostic workflows, and, in some contexts, prediction of molecular alterations.
A common theme is breadth: cell-level features, tissue architecture, slide-level context, and case-level summaries can coexist in one product.
Handling real-world variation
Variability across laboratories is unavoidable: scanners differ, staining protocols drift, pre-analytic conditions vary, and case mix changes by site. Earlier task-specific models were often trained on limited datasets, making generalization difficult.
Because foundation models are trained on large, heterogeneous data, spanning sites, geographies, and technical variation, they are more likely to learn robust features. The foundation model effectively compresses raw images (including artifacts) into embeddings that capture clinically relevant morphology while reducing sensitivity to nuisance variation. Downstream task models can then focus more on biology and less on noise.
As digital pathology adoption expands, the scale and diversity of available WSIs should increase substantially. That growth can further improve foundation models and, by extension, raise the baseline performance of many downstream tasks.
Validation breeds trust
Everything we know about quality in the laboratory applies to AI. For any model, it is essential to understand how it was validated: the cohort, the sites, the scanners, the staining conditions, and the limitations.
For many pathology AI tools, a practical advantage is that the pathologist still has the original image. Model outputs are interpreted in context, often with visual overlays such as heatmaps, alongside clinical history and specimen type. That combination of rigorous validation and informed human oversight enables trust.
Evolution comes with experience
As these tools mature, the role of the pathologist and the laboratory will evolve, not away from expertise, but toward new capabilities and more efficient workflows. For labs beginning a digital transition, this is a favorable moment: infrastructure is more reliable, scanning quality is higher, and the product ecosystem is more mature.
My advice is straightforward: explore what is available. Do demos, speak with peer laboratories, and build experience with tools that have been clinically validated. The gap between what was only theoretically possible a few years ago and what is practically available today is substantial.
Better models benefit patients
The goal is aligned with the laboratory’s mission: getting the right diagnosis to the right patient as quickly and reliably as possible. Foundation-model-driven products can help deliver faster time to diagnosis, greater consistency, and better access to quantitative biomarkers that support precision medicine.
Looking ahead, as digital pathology becomes more widespread, patients may gain access not only to their digital slide images but also to AI-generated insights about their disease. A more transparent diagnostic journey could help patients become more informed partners in their care.
Architecting the future
Beyond the clinical lab, foundation models are enabling new capabilities in translational research ‒ such as the discovery of novel tissue pathology-based predictors of treatment response ‒ and supporting the development of drugs and companion diagnostics. One encouraging example from our own experience was training and deploying PathAI’s largest foundation model to date. After the switch to this updated foundation model, we observed a more than 10 percent relative improvement across a diverse set of tasks, from cohort-level outcome prediction down to micron-scale classification of diverse cell types. The common driver was richer embeddings that transferred broadly across tasks.
The compounding effect is the real opportunity. As foundation models improve through larger, more diverse data and better training methods, the task-specific models built on top of them tend to improve as well. The remaining work is practical and implementation-driven: integrating these tools into real workflows, learning quickly from pathologists’ feedback, and tuning products to deliver reliable, measurable value in clinical practice.
With that feedback loop in place, we expect meaningful gains to continue, helping the field move beyond a handful of narrow models toward a durable portfolio of foundation-model-enabled products that support pathologists and, ultimately, patients.
Newsletters
Receive the latest pathologist news, personalities, education, and career development – weekly to your inbox.
