Why Computational Pathology Needs a Different Kind of AI Team

I’ve spent the last fifteen years watching computational pathology from almost every seat in the room — as a pathology resident at the NIH, as a pathologist at Bristol Myers Squibb, as the first pathologist Alphabet hired (clinical lead on the AR microscope and the Gleason grading work at Google), as Head of Pathology Data Science at AstraZeneca, and as founding Chair of the Division of Computational Pathology & AI at Mayo.

Across every one of those seats, I kept running into the same problem. And it wasn’t the algorithm.

The bottleneck isn’t the model

Most computational pathology work that stalls doesn’t stall because the model is wrong. It stalls because the team around the model is structured wrong.

Large academic labs have the clinical expertise but move on grant cycles. Tech companies move fast but don’t have enough pathologists in the room to ask the right question in the first place. Biopharma has the trial data but ships on regulatory timelines that can’t absorb a rapid learning loop. Every structure I worked inside had a version of this — different on the surface, identical underneath.

The result is a field that over-produces publications and under-produces things that change what happens to a patient.

What actually moves the needle

What I’ve seen work is simpler than the current org charts suggest: the pathologist defining the question is sitting next to the engineer writing the training loop. The designer is in the same thread. The feedback loop from “this is interesting” to “let’s run it on a different cohort” is days, not quarters.

That is not a new insight. Anyone who has built something useful in this field already knows it. The hard part is building an organization where that is the default — not a temporary arrangement that survives until the next re-org.

What we’re building at Cura Labs

Cura Labs is deliberately small. Not because we’re just starting — because the overhead of coordination is the single biggest tax on research velocity, and after fifteen years of paying that tax inside much bigger organizations, I wanted to build somewhere it doesn’t exist.

A few things fall out of that constraint:

• Hypothesis to analysis in days. When I see a pattern worth investigating, Ari can have a pipeline running on the dataset within 48 hours. No grant. No committee. No six-month procurement cycle for compute.
• AI as infrastructure, not just a model. A trained model is one piece of the work. The annotation, comparison, visualization, and reporting around it — the thing a pathologist actually touches — is the real product.
• Clinical grounding on day one. Every project starts with a question a pathologist would actually ask. Not can we classify this, but does this classification change what happens to the patient. That one filter eliminates a lot of work that looks good in a paper and changes nothing in the clinic.

Where this is going

The data is getting better. Whole-slide imaging is becoming standard. Foundation models trained on histopathology are starting to work. The missing piece is not the technology — it’s the team shape that can turn the technology into a discovery that matters.

The teams that will do the most for patients over the next five years, in my view, will not be the largest. They will be the ones where the pathologist and the engineer never had a handoff in the first place.

That’s the team I’ve wanted to be on for fifteen years. It’s the team we’re building now.