AI in pathology is not a single challenge: what toxicology is teaching us

Here, the goal is to systematically assess for any adverse effect, including those that may be rare, subtle, or previously unseen. As Salma Dammak (Bigpicture member in work package 4, and Postdoctoral Researcher in AI for Toxicologic Pathology at Radboud University Medical Center) explains, many of these challenges are “very much needle in a haystack” problems.

That shift, from answering narrowly defined questions to searching for the unexpected, changes how AI needs to be developed, validated, and ultimately used. It is also where important lessons for the broader field begin to emerge.

From narrow questions to open search

The distinction between clinical and toxicologic pathology is not about different technologies, but different questions.

As Santiago Villalba (Bigpicture EFPIA lead for work package 4, and Machine Learning Researcher at Bayer) describes it, clinical pathology typically focuses on a specific patient and a relatively well-defined task. Toxicologic pathology, by contrast, asks a broader question: is a candidate therapeutic molecule safe? Currently this is studied in controlled animal studies across many organs, doses, and species. That difference matters.

In toxicology studies, pathologists are therefore responsible for identifying any adverse effect, known or unknown, and mechanistically understanding its biological meaning. This makes the problem inherently more open-ended. It also means AI systems need to be developed to  make efficient use of scarce information over a wide range of possible toxicological findings and tissue types.

When data does not look the way AI expects

This broader scope directly impacts how data can be used. In many toxicology settings, annotations are limited. Labels exist in large quantities as they are captured in routine, but only at the specimen level for each individual, rather than pinpointing exact locations on a slide. At the same time, abnormalities are rare compared to the large volume of normal tissue. “We really have to be careful around how we design the models”,  Salma notes, “to make sure that when there are cases that are unexpected, the model behaves in a way that is how we would like it to.”

As a result, approaches such as weakly supervised and self-supervised learning become essential. Rather than relying on detailed annotations, these methods allow models to learn from broader patterns in the data. In one example publication that addressed this, Salma and her team present a model which was trained only on normal tissue and then used to detect abnormalities by measuring how much a new sample deviates from that learned baseline. The result is not a perfect solution, but it reflects a broader shift. In toxicologic pathology, AI must be designed not only to recognize what is known or already observed in the data, but to remain robust when encountering what is not.

More than performance: trust, robustness and workflow

Despite rapid progress in AI development, routine use in toxicologic pathology remains limited. This is not simply a question of model performance. As Santiago highlights, “one of the key challenges in all of digital pathology” is the lack of detailed supervision to train AI models, but even that is only part of the picture. Models must also generalize across different conditions, like different scanners and experimental protocols. To ensure adoption,  these models must fit into fast, highly ergonomic workflows. And they must earn the trust of pathologists working in regulated environments.

Explainability and uncertainty estimation are therefore not optional features. They are central to adoption. AI systems need to show not only what they predict, but why, and how confident they are in those predictions. Without that, even strong performance metrics are not enough to support real-world use. In a preprint, Santiago and his team show that interrogating which parts of the slides the AI model uses may be different from what we expect, for example focusing on a lesion’s edges and its surroundings rather than its center. This shows that AI model attention may be different than we expect and may even highlight potential areas of exploration that we may have not considered for this task.

Where progress is already happening

While full integration into routine studies is still developing, AI is already contributing in more targeted ways. Applications such as biomarker quantification, quality control, and early research-stage decision support are beginning to show value. These use cases often sit outside highly regulated workflows, allowing organizations to explore the benefits of AI while building experience and confidence.

At the same time, the field is moving toward more generalizable approaches, including foundation models trained on large collections of unlabeled data. These models have the potential to support a wide range of downstream tasks with higher accuracy and lower data requirements, from lesion detection to cross-domain applications. This is where scale and diversity of data become essential.

What this means for Bigpicture

One of the structural challenges in pathology AI today is access to sufficiently large and diverse datasets, particularly in the non-clinical domain. Bigpicture is helping address that gap.

By bringing together data across species, organs, and study types, capturing a large spectrum of technical and biological variability, and by enabling collaboration between pharma, academia and clinical partners, the project is creating the conditions for more robust model development and meaningful benchmarking.

But its role goes beyond data alone. The collaboration itself is part of the value. Pharma contributes large-scale preclinical datasets and applied questions. Academia and hospitals bring methodological depth, clinical insight, and longer-term research perspectives. Together, this creates a more complete environment for advancing AI in pathology.

It also opens the door to a broader ambition: strengthening the connection between preclinical and clinical research. As Salma says, “the dream too is being able to connect, from the preclinical to the clinical, and back.”

From insight to implementation

The direction of travel is clear. AI in pathology is advancing rapidly, and the technical foundations are strengthening. At the same time, the next phase will depend on more than algorithms. Implementation, validation, workflow integration, and regulatory alignment will shape how quickly these tools become part of everyday practice. Demonstrating practical value, through concrete use cases and measurable impact, will be critical. As Santiago reflects, progress will require a shared effort: “We need to do a lot of work, all of us, and have a common vision. It is truly exciting.”

A broader lesson for pathology

Toxicologic pathology may represent one of the more complex applications of AI in the field. But that complexity is also instructive. It highlights the limits of narrow, task-specific approaches. It underscores the importance of robustness, adaptability, and trust. And it shows that advancing AI in pathology is not only about improving models, but about building the right data, workflows, and collaborations around them.

In that sense, the lessons from toxicology extend well beyond the field itself. They point toward a broader shift in how AI in pathology is being understood, and how it may ultimately be used.