The Iron Man Suit for Radiologists: Epsilon Health Exits Stealth With $27.6M and 1% of America's Daily X-Rays

On September 10, 2026, a San Francisco startup called Epsilon Health emerged from stealth with $27.6 million in funding and a claim that would have sounded absurd a few years ago: in less than ten months of operation, its AI-native radiology practice is already on track to interpret roughly 1% of all daily X-rays taken in the United States.

The round was led by AlleyCorp, with participation from Uncork Capital, Renegade Partners, SemperVirens, and Jack Altman. Founded in 2024 by CEO Rustin Rassoli — who grew up in the radiology industry and has said he believes “if he doesn’t build this, nobody will” — Epsilon is not building another AI tool for radiologists to bolt onto their workstations. It is building a new kind of radiology practice, from the ground up, with AI woven into the clinical workflow itself.

A System That Cannot Be Fixed With More Software

The U.S. radiology industry is in a quiet crisis. More radiologists are retiring than entering the field, while demand for imaging keeps climbing. Interpretation times have tripled since 2014. Patients can get an X-ray in minutes, but wait days for someone to read it — and for a patient with a potentially serious condition, those days matter.

The industry’s answer so far has been AI tools — more than 750 of them have been built for radiology, backed by billions of dollars in investment. The results have been underwhelming: despite the abundance of tools, fewer than one in three radiologists actually use them, and wait times are now longer than they were a decade ago. As Uncork Capital’s Tripp Jones put it in the firm’s investment memo, “The system is broken, and you can’t fix it by selling more software.”

The arithmetic of the problem is stark. The average radiologist reads about 50 studies a day — alongside consultations, protocols, and everything else that fills a clinical day. The bottleneck isn’t image acquisition; modern scanners produce images in minutes. The bottleneck is interpretation capacity, and it cannot be scaled by hiring, because the talent pipeline is shrinking.

Not a Tool — A Practice

Epsilon’s answer is structural rather than incremental. Instead of shipping another piece of software that radiologists have to learn, the company operates what it calls an AI-native radiology practice: AI, physician radiologists, and the clinical workflow combined into a single system designed to make each radiologist dramatically more capable.

The numbers the company reports are significant. Scans are interpreted 2-3x faster than before, which effectively doubles or triples the capacity of the system. Each radiologist becomes far more productive without having to change the way they work — no new interfaces to master, no workflows to rebuild. Jones describes the result as “the Iron Man suit for radiologists.”

That framing matters, because it distinguishes Epsilon from the two dominant AI-in-radiology models. The first is the “software vendor” model: sell a tool to hospitals and hope it gets adopted into workflows that were never designed for it — the model that has left three-quarters of radiologists not using AI at all. The second is the “autonomous AI” model: algorithms reading scans without physicians, which remains a regulatory and liability minefield in the U.S. and faces deep institutional resistance. Epsilon occupies a third position — an AI-native practice that contracts with radiologists and imaging providers, keeps physicians firmly in the loop, and re-architects the entire interpretation pipeline around the human-machine combination.

The Data Flywheel

The most strategically interesting part of Epsilon’s model is its compounding data advantage. The company’s AI improves every time a radiologist uses it, because every correction a physician makes becomes new clinical training data. Unlike models trained on textbook datasets or labeled data bought off the shelf, Epsilon’s models are refined on real doctors correcting real cases in the course of their daily work.

This creates a classic flywheel: more studies read → more physician corrections → better models → faster and more accurate reads → more customers → more studies read. It is the same structural advantage that has played out in other AI-native service businesses, and it is difficult for incumbents to replicate because the data comes from the operation of the practice itself.

The market has noticed. Some of the largest imaging providers in the country have evaluated what Epsilon built and concluded it was easier to become a customer than to recreate it in-house — a signal that the company interprets as validation that the advantage is real.

The Traction

For a company that has been operating for less than a year, Epsilon’s scale is unusual:

• More than 250,000 patients served since launch
• Roughly 2,500 studies processed per day
• On track to interpret 1% of all daily U.S. X-rays in its first year

That last number deserves context. The United States performs on the order of hundreds of thousands of X-ray studies daily across hospitals, clinics, and imaging centers. For a startup founded in 2024 to be reading a full percentage point of that volume — not as a pilot, but as production clinical work — suggests the model is not a demo. It is load-bearing infrastructure for the customers who depend on it.

The team behind the company explains some of the speed. Chief Medical Officer Dr. Roi Bittane previously ran the largest radiology group in the country and still reads images almost every day. Arjun Karpur, who leads the ML and AI team, is a former DeepMind researcher. Founder Rustin Rassoli grew up around the industry and has watched the shortage crisis unfold for years, saying he saw “firsthand what growing backlogs and radiologist shortages do to patients who are just waiting for an answer.”

Why This Model, Why Now

Epsilon’s launch lands at a moment when the debate over AI in medicine has largely polarized between hype and fear. This company is a useful data point in that debate because it is neither a promise nor a threat — it is an operating business with real clinical volume.

The AI-native practice model also sidesteps two problems that have stalled AI adoption in radiology. First, the adoption problem: hospitals buy tools that radiologists then ignore, because the tools add friction to an already overloaded workflow. Epsilon’s customers don’t face that problem because Epsilon owns the workflow. Second, the liability problem: autonomous AI reading scans without physician sign-off remains contentious. In Epsilon’s model, physicians interpret with AI amplification — the “Iron Man suit” — which fits comfortably inside existing medical-legal structures.

The risks are real too. Radiology groups and hospitals may resist outsourcing interpretation to a new practice; incumbent teleradiology companies could replicate the AI-native structure with their own capital; and the regulatory environment for AI-assisted interpretation could tighten as volume grows. Scaling a clinical practice is operationally harder than scaling software, and quality must hold as volume multiplies.

But the underlying math favors whoever solves this. Demand for imaging rises every year, the radiologist pipeline shrinks every year, and 750+ point solutions have failed to bend the curve. Interpretation capacity is the constraint, and the only credible way to expand it is to make each radiologist dramatically more productive. Epsilon’s early numbers — 2-3x faster reads, 2,500 studies a day, 1% of U.S. X-rays in year one — are the first evidence that someone may have actually done it.

For patients, the stakes are simple. The X-ray that takes minutes to capture should not take days to read. A startup from San Francisco just raised $27.6 million on the bet that it can close that gap — and it is already reading your scans.