Healthcare IT · Medical data annotation
Centaur Labs
A Boston startup that pays medical students and clinicians to label medical images, sounds and signals inside a mobile app, then combines their answers statistically to beat any one expert. Peer reviewed papers now test that claim.
A wife's flashcards
Erik Duhaime was finishing a PhD at MIT's Center for Collective Intelligence when he noticed his wife, then a medical student, spending hours on flashcard and quiz apps. His own research had already shown something odd: a group of medical students, if measured for accuracy and combined at only their best, could out-diagnose professional dermatologists reading the same skin lesion images.
“I realized my wife’s studying could be productive work for AI developers,” Duhaime told MIT News in 2023. He built DiagnosUs, a mobile app where people review real medical images, audio clips and other clinical data, get scored against known answers, and are paid small cash prizes when they perform well. The pooled, skill weighted opinions become training labels for medical AI companies.
The problem it targets
Healthcare generates close to a third of the world's data by some industry estimates, and most of it is unstructured or poorly labeled. Medical AI models need large volumes of accurately labeled images, audio and text, and the people qualified to label a chest CT or an EEG trace are the same radiologists and neurologists already in short supply. Outsourced labeling firms use generalist workers who lack that training; hiring specialists directly is slow and expensive.
Centaur's pitch is that a large, measured crowd of semi experts, mostly medical students and practicing clinicians, can match or beat a small panel of full experts if accuracy is tracked and the best answers weighted more heavily. “What interested me was the wisdom of crowds phenomenon,” Duhaime said.
Founders and early support
Duhaime founded Centaur with Zach Rausnitz, a longtime friend from Brown University who became chief technology officer, and Tom Gellatly, who had managed the data labeling engineering team at the self driving car company Cruise Automation and previously led mobile development at Sidecar. Duhaime completed his doctorate under Thomas Malone, founding director of MIT's Center for Collective Intelligence.
The company drew on MIT's own entrepreneurship pipeline: a grant from MIT's Sandbox Innovation Fund in 2017, the delta v summer accelerator in 2018, and a spot in Y Combinator's Winter 2019 batch. SEC records list Centaur Diagnostics, Inc. as incorporated in Delaware in 2017, matching the Sandbox grant year.
From seed to Series B
Centaur raised a $15 million Series A in September 2021, led by Matrix Partners with Accel, Global Founders Capital, Susa Ventures, Y Combinator and individual investors including WHOOP founder John Capodilupo, One Medical founder Tom Lee and PillPack founder Elliot Cohen. A Form D filed with the SEC on August 17, 2021 puts the actual amount at $15,885,849 sold to 38 investors, slightly above the round's headline figure.
Three years later, in October 2024, Centaur announced an oversubscribed Series B of more than $16 million led by SignalFire, joined by new investors Samsung Next and Alumni Ventures alongside existing backers Matrix, Accel, Susa, Omega and Y Combinator. The announcement introduced an on demand labeling product promising expert annotations inside an hour rather than days. SignalFire's own blog post about the deal, published the same week, independently confirms the round and a board seat for SignalFire executive in residence Jeff Butler, the former CEO of Privia Health.
Widening beyond healthcare
Centaur's original pitch was medical data exclusively, and its strongest evidence still comes from healthcare: partnerships with researchers at Brigham and Women's Hospital, Massachusetts General Hospital and Memorial Sloan Kettering have produced peer reviewed papers on lung ultrasound, ICU EEG and skin lesion labeling. But the company's own blog in 2025 began pitching the same approach for content moderation, climate data, supply chain documents and robotics, alongside its continuing health work. In 2026 the company also uses its Centaur.AI brand more than Centaur Labs, including in a Nature Medicine paper co-authored by Duhaime.
What is proven, and what is still claimed
| Evidence | What the record shows | Source type |
|---|---|---|
| Lung ultrasound labeling | With Brigham and Women's Hospital researchers, crowdsourced DiagnosUs labels for B-line classification on 198 test clips matched trained experts: 87.9% crowd concordance versus 85.0% expert concordance (P=.15, not significant). Against a standard excluding each expert's own vote, the crowd beat the average expert, 87.4% versus 80.8% (P<.001). | Peer reviewed |
| ICU EEG labeling | A DiagnosUs scoring contest found weighted crowd votes non-inferior to 8 named experts on seizures and periodic patterns: crowd accuracy .70 versus expert accuracy .68 across 478,834 answers from 1,542 participants. No individual expert outperformed the crowd. | Peer reviewed |
| Skin lesion labeling | Wisdom of crowd algorithms tested on International Skin Lesion Challenge 2018 images, labeled by app recruited annotators, outperformed expert benchmarks in the published results, with U.S. government research support. | Peer reviewed |
| Bias correction research | A 2026 paper with Centaur Labs and Indiana University co-authors found that recalibrating crowd workers' probability judgments on DiagnosUs reduced bias and improved downstream model accuracy across two experiments. | Peer reviewed |
| Explainable AI research | Erik Duhaime, credited to Centaur.AI, co-authored a 2026 Nature Medicine paper with MIT, Columbia and Stanford researchers on AI explanations and diagnostic decisions. It studies human-AI interaction generally, not Centaur's own labeling product. | Peer reviewed |
| Federal grant | NIH grant 1R41LM015320-01, $312,528, to “Centaur Diagnostics, Inc.” (dba Centaur.AI), PI Erik Duhaime, for a Sep 2026 to Aug 2027 project on free text annotation with Vanderbilt University. | Public record |
| Named customers | Massachusetts General Hospital, Memorial Sloan Kettering, Eight Sleep, Elsevier's Scibite, Activ Surgical and Medtronic are named in the Series B release. Zeta Surgical's CTO and a Paige AI scientist are quoted by name in a SignalFire post, not Centaur's own materials. | Company and partner stated |
| Scale and ROI figures | The company reports customers seeing “up to 20X ROI from annotation speed improvement,” a network of “over 50,000 experts,” and an Eight Sleep snore model improved from 70% to 93% accuracy. None of these figures have been independently audited. | Company-stated |
Read plainly, Centaur's core scientific claim, that a measured and weighted crowd of semi experts can match or beat individual specialists on narrow labeling tasks, has peer reviewed support across lung ultrasound, EEG and dermatology datasets, mostly from studies co-authored with hospital researchers rather than written by Centaur alone. The business claims on top of that, growth figures, ROI multiples and the roster of enterprise customers, come from the company and its investors and have not been independently audited.
What to watch
- Whether a Form D or amendment appears for the October 2024 Series B; as of September 2026 the only SEC filing on record for Centaur Diagnostics, Inc. is the 2021 Series A notice.
- Results from the new NIH funded free text rationale project with Vanderbilt, running through August 2027.
- Whether Centaur's push into non-healthcare labeling dilutes or strengthens the medical specialization its published research is built on.
- Named, published case studies from the enterprise customers the company lists, beyond the two SignalFire-sourced quotes.
In their words
“I realized my wife’s studying could be productive work for AI developers.”
Erik Duhaime, co-founder and CEO, to MIT News, 2023 · Independent
“Today we have tens of thousands of people using our app, and about half are medical students who are blown away that they win money in the process of studying. So, we have this gamified platform where people are competing with each other to train data and winning money if they’re good and improving their skills at the same time.”
Erik Duhaime, to MIT News, 2023 · Independent
“What interested me was the wisdom of crowds phenomenon.”
Erik Duhaime, to MIT News, 2023 · Independent
“It is difficult to curate large medical datasets, and nearly impossible to source accurate labels from those with medical knowledge and specialized training.”
Erik Duhaime, in the Series A funding announcement, 2021 · Company release
“In healthcare, where AI hallucinations can cost lives, ‘garbage in, garbage out’ data problems are unacceptable and models need to be ongoingly evaluated and monitored once they’re deployed.”
Erik Duhaime, in the Series B funding announcement, October 2024 · Company release
“Centaur tackled this problem quickly and accurately, enabling us to focus our time on building cutting-edge AI.”
Raahil Sha, CTO, Zeta Surgical, quoted by SignalFire, October 2024 · Customer, independent
“Centaur Labs annotated the first batch of data so quickly it seemed unreal.”
Fausto Milletarì, Senior AI Scientist, Paige, quoted by SignalFire, October 2024 · Customer, independent
“There were times in the clinic where I realized that I was doing better than others because of what I learned on the DiagnosUs app.”
Andrews Gyabaah, medical student and DiagnosUs contest winner, to MIT News, 2023 · Independent
Related companies
Sources
- Public recordForm D, Centaur Diagnostics, Inc.
- Public recordNIH RePORTER award 1R41LM015320-01
- Peer reviewedGamified Crowdsourcing as a Novel Approach to Lung Ultrasound Data Set Labeling
- Peer reviewedEvaluating crowdsourcing for ICU EEG annotation: A comparison with expert performance
- Peer reviewedBoosting wisdom of the crowd for medical image annotation using training performance and task features
- Peer reviewedImproving crowdsourcing for AI through cognitive-inspired data engineering
- Peer reviewedDivergent impacts of explainable AI for dermatological diagnosis on clinicians versus lay people
- IndependentGamifying medical data labeling to advance AI
- IndependentCentaur Labs Gets $15 Million To Improve Data For Healthcare AI
- IndependentCentaur Labs raises $15 million, led by Matrix Partners
- Investor, independentCentaur Labs: Transforming medical data annotation
- IndependentCentaur Labs company page
- CompanyCentaur AI Series A funding announcement
- CompanyCentaur AI raises $16M Series B to accelerate AI for health and science
- CompanyCentaur.ai website and news page
Profile researched and written by Healthcare Discovery. Last updated September 29, 2026.
