Nature Medicine · 2019350+ citations
Augmented-reality microscope with real-time AI for cancer diagnosis
First clinical demonstration of a real-time AI overlay during pathologist diagnosis. Co-invented by Dr. Hipp.
Frontier projectCoa · Digital Pathology
We’re embarking on one of our most ambitious projects: end-to-end computational pathology.
A whole-slide viewer, a frontier-model benchmark, and governed clinical AI — led by a pathologist who helped invent the field.
Research imagery — no patient data.
- Fifteen years
- Building pathology AI — NIH to Google to in-house
- 40+
- Peer-reviewed articles
- 3
- Anchor papers that defined the field
- 1st
- Pathologist hired at Google Health
The problem
Every diagnosis runs through a pathologist. There aren’t enough of them.
The answer comes from the slide, and the slide comes to a pathologist already drowning in volume. Demand climbs every year as the workforce shrinks.
-
The workforce is shrinking
US pathologists fell about 17% in a decade. The ones who remain carry more.
-
Volume is exploding
More cases, harder cases, rising cancer incidence — demand outruns capacity.
-
Turnaround drags
Slides queue and sign-out waits, so the answer a patient needs arrives late.
-
Access is unequal
Whole regions have almost no pathologists — roughly one per million in places.
See our sources
- Metter et al., JAMA Network Open · Trends in the US and Canadian Pathologist Workforces, 2007–2017 — the active US pathologist workforce fell about 17.5% over the decade, even as workload per pathologist rose.
- Robboy et al., Archives of Pathology & Laboratory Medicine · Pathologist Workforce in the United States — projected a widening gap as case volume and complexity climb while the workforce ages and shrinks.
- Wilson et al., The Lancet · Access to Pathology and Laboratory Medicine Services — access is critically scarce in much of the world — parts of sub-Saharan Africa have roughly one pathologist per million people.
- IARC / WHO, GLOBOCAN · Global Cancer Statistics & 2050 Projections — new cancer cases are projected to rise sharply through 2050, driving diagnostic demand past current capacity.
How we answer it
We’re building the read, end to end.
Three builds, each answering a part of the process — see the slide, trust the model, reason beyond it.
01See the slide
The whole-slide viewer your team reads on every day.
Browser-native deep-zoom, case routing, and an integration surface for FDA-cleared AI — built to sit inside the sign-out a pathologist already runs.
Browser-native
Deep-zoom whole slides in the browser — no install, no plugin, no workstation.
Case routing
The worklist triaged and routed, so the next slide up is the right one.
An AI surface
A place for FDA-cleared models to overlay their read, grounded to the pixels.
FHIR-native
Results land in the LIS and EHR as standard resources, not a separate tool.
Representative interface — product screenshots to follow.
AI belongs in the eyepiece, not the pathologist’s chair. It flags the focus and shows its work — the sign-out stays a human decision.
02Trust the model
We test the models. We don’t just use them.
A multi-model evaluation viewer runs the same region past the frontier models side by side and scores them against known-answer cases — which reads this stain best, at what confidence, at what cost.
Side by side
The same region, the same question, every frontier model at once.
Scored, not vibed
Graded against known-answer cases, with agreement and confidence on the record.
Cost-aware
Quality and cost per read together, so the right model runs the right task.
The gate
The discipline behind every model that reaches a clinical surface.
Same region · four models graded vs. known answer
Model A0.96agrees
Model B0.91agrees
Model C0.83partial
In-house0.88agrees
Consensus focus · concordant with ground truthcost / read shown
Representative interface — model labels and scores illustrative.
03Reason beyond it · Coming soon
PathLens — pathology reasoning, beyond the slide.
Pathology-grade interpretation for general medicine: records, labs, and wearable signal synthesized into one read, so there are fewer touchpoints between a question and an answer.
Multi-modal
Records, labs, and wearable signal, read together instead of one at a time.
Built on Oracle Health
On the record layer clinicians already run, not a data island beside it.
PathLens · unified readrecords · labs · wearable
Records Labs Wearable Pathology
Interpretation
Concept interface — in development with clinical partners.
Built for clinical trust
In medicine, the read has to be governed.
-
Governed by design
Grounded, measured, auditable. AI proposes; the pathologist decides. Mapped to NIST AI RMF and OWASP for LLMs.
-
Interoperable · FHIR
HL7 FHIR into the LIS and EHR, so a read lands in the workflow, not beside it.
-
Private · HIPAA
PHI stays inside the tenant. Access scoped and logged; no case content trains a shared model.
-
FDA-cleared surface
The viewer hosts FDA-cleared AI, assistive by design, with the pathologist at sign-out.
Led by a pathologist who helped invent the field
Dr. Jason Hipp, MD, PhD.
The first pathologist hired onto Google Health’s computational-pathology team, co-inventor of Google’s AR Microscope and its SMILY image search, and founder of a hospital Division of Computational Pathology & AI — the rigor he brought to Mayo, Alphabet, and AstraZeneca, pointed at the slide.
- Alphabet / Verily · Google Health — first pathologist hired; co-inventor of the AR Microscope and SMILY (US Patent 11,379,516).
- Mayo Clinic — founder, Division of Computational Pathology & AI.
- AstraZeneca — Executive Director, Pathology Data Science & Innovation.
- Federal training — NIH Clinical Center, Walter Reed, and the Armed Forces Institute of Pathology.
MD/PhD · Board-certified Anatomic Pathology · Wake Forest
Published, patented, and FDA-recognized — three papers that helped define the field
npj Digital Medicine · 2019570+ citations
Deep learning for Gleason scoring of prostate-cancer biopsies
A foundational paper for clinical-grade prostate-cancer pathology AI.
J. of Medical Imaging · 2014FDA + NIH coauthors
An evaluation environment for digital and analog pathology
A platform for the validation studies clinical AI has to pass — a decade before validation was common.
The methodology generalizes beyond the slide. See clinician-directed AI and nine years of wearable data — a diagnosis in days that four years of visits had missed.