Artificial intelligence has advanced virtually every industry, including healthcare. But AI is only as accurate as the data it’s trained on — meaning that if it isn’t properly labeled and programmed, it can be pretty unreliable, and even dangerous. Training AI algorithms for healthcare is quite a process, requiring massive data sets of medical images, videos, text and audio recordings, all of which have to be accurately labeled by humans. It’s a big job.
Here to tackle this challenge is Centaur Labs, a Boston tech startup that has cultivated a network of tens of thousands of people to do the labeling. The company announced Friday it raised a fresh $15 million in Series A funding. The round was led by Matrix Partners, with participation from familiar faces like Accel, Y Combinator and John Capodilupo, co-founder and CTO of local standout Whoop.
“Centaur’s technology doesn’t just offer data labels, it rethinks the medical second opinion and harnesses a network of trusted experts,” Stan Reiss, a partner at Matrix, said in a statement. “The ability for AI to make an impact in healthcare depends on the ability to solve the data labeling bottleneck, and Centaur will catalyze the development and adoption of AI solutions throughout the industry.”
‘That’s How the AI Learns’
Erik Duhaime, the co-founder and CEO of Centaur Labs, likens the company’s model to those “I’m not a robot” tests Google sometimes makes you do — the ones where you have to identify trees, or cars, or traffic lights. While these exercises are good at proving you’re human, they’re also good at helping Google annotate data for self-driving cars.
“Just like all AI these days, [self-driving cars] need millions and millions of images showing these are street signs, these are cars, these are trees, etc. That’s how the AI learns,” Duhaime told Built In. “The same is true for healthcare AI.”
Of course, you can’t expect random strangers on the internet to be able to identify a lung tumor or a heart attack, so medical data labeling looks a little different. Ordinarily, it’s done by hiring expert physicians at expensive hourly rates to annotate cases, which Duhaime says is often very slow and ends with these experts disagreeing with each other anyway. Instead, Centaur Labs has built up a global network of thousands of medical students and professionals to do the annotating, and gamifies the experience with things like cash prizes and a leaderboard to encourage correct identifications.
“Even the world experts will have a bad day, or will be doing labeling when they’ve had one too many glasses of wine. So because we’re continually measuring the accuracy of people, we know whether they’re doing a good or a bad job at the task, which helps us know when to reward them,” Duhaime said, adding that getting input from multiple people will inevitably lead to better accuracy.
“A person plus a computer is going to do better than either alone,” Duhaime continued. “Even better than one individual or one algorithm with each other, is going to be a network of people and computers who are trusted based on performance metrics to solve medical data annotation problems.”
‘Picks and Shovels’ of the ‘Medical AI Gold Rush’
Most of these students and professionals do this work on Centaur Labs’ app, called DiagnosUs — “it’s, quite literally, a gamified, competitive, medical data annotation app,” Duhaime explained. Users simply sign on and start identifying medical images. If they do well, they gain ranking on a leaderboard, and can win money if they outperform others.
This model is a win-win. Obviously, healthcare AI algorithms continue learning, which helps the industry progress. But the users, especially the medical students, get something out of it too. For one, cash (although Duhaime says it’s mostly “coffee money”), but also the ability to sharpen their skills. Instead of having to pay hundreds of dollars for flashcard apps and various studying tools, they’re signing onto DiagnosUs and “earning money for something that many of them would be willing to pay for,” Duhaime said.
Long term, he added, this network of medical experts and students could do more than just improve the accuracy of existing medical AI. It’s a continual process, where new hardware and new patient populations will need to be addressed. He also anticipates that Centaur Labs will eventually be more integrated with the diagnostics process, envisioning a direct-to-consumer offering where a person can send in a chest X-ray and the company’s network of vetted annotators will identify if they have a collapsed lung.
“That’s not our focus right now, but it speaks to our long-term grand vision,” Duhaime said. “All we want to do is be the picks and shovels for the medical AI gold rush.”
A surge of VC funding has, indeed, poured into this space lately. Earlier this year, AI-based digital diagnostics startup Paige raised a $100 million Series C round; then Boston-based PathAI, which uses AI to advance disease research and drug development, raised $165 million in Series C funding. Just last month, Iterative Scopes, an MIT spinout that uses AI to analyze gastroenterology images, got a $30 million investment from pharmaceutical giants Eli Lilly and Johnson & Johnson.
Because of the pandemic, it appears the medical industry is beginning to embrace AI and its potential, which, of course, means that ensuring its accuracy and reliability is more important than ever. This, plus the normalization of a remote and mobile-first workforce over the last 18 months, leaves Centaur Labs primed for success going forward.
To keep up the momentum, the company will use this fresh funding to expand its global network of annotators, accelerate product development and grow its team, with a focus on engineering, operations, and business development roles. Duhaime says Centaur Labs now has about a dozen employees at its office here in Boston, and that the team will likely double in the next 18 months.
