Medical AI · Radiology reporting
Mecha Health
Four University College London researchers built a chest x-ray model that writes the whole report, published benchmarks saying it matches or beats radiologists, and raised $4.1 million. In the United States a report generator is a regulated device, and no FDA authorization for the company is on record.
A thousand x-rays, and nothing else
In June 2025 Mecha Health published a test it designed itself. It pulled 1,000 chest radiographs at random from the MIMIC-CXR test set, a public archive of images paired with the reports radiologists wrote for them, and gave those images to board certified radiologists at US attending or UK consultant level.
The humans got the picture and nothing else. No patient history. No prior studies. No lateral view. The company's model, Mecha Net v0.2, was given the same bare input so the comparison would be fair.
Then both sets of reports were scored against the original ones on eight measures, from word overlap to CheXbert, a labeller that reads a report and extracts fourteen clinically important findings. The company reported that its model came out ahead of the radiologists on every measure.
Its own conclusion was careful in wording and large in scope: “To the best of our knowledge, this is the first work to evidence that radiology foundation models can systematically match or exceed human performance for the task of report generation by analysing medical images.”
The study was run and published by the company selling the model, it has not been peer reviewed, and the radiologists were reading with one hand tied, a condition no real radiologist works in.
The gap it is aimed at
The demand side is not in dispute. In its own note on diagnostic error the company quotes the 2015 National Academies finding that “Most people will experience at least one diagnostic error in their lifetime, sometimes with devastating consequences.” It cites daily radiology error rates of 3 to 5 percent, retrospective rates as high as 30 percent, and a study in which 19 percent of lung cancers with a median diameter of 16 mm were missed on chest x-ray.
The company's funding announcement adds the workforce numbers it uses in every pitch: nearly seven in ten radiologists reporting short staffing, 67 percent saying burnout reaches their personal lives, imaging demand per patient up 13 percent, training capacity nearly flat for decades.
Chest radiography is the busiest imaging exam there is, about 70 million US exams a year by the reckoning of the competitor Qure.ai. Most commercial AI narrows that to one finding at a time. Mecha Health went the other way: one generalist system, a whole report, priced per scan.
Two doctors, a mathematician and a surgeon's engineer
Ahmed Abdulaal trained in medicine at Imperial College London and became a Microsoft PhD scholar in machine learning at UCL, with a stint as a research scientist at AstraZeneca's Centre for Artificial Intelligence. Nina Montaña Brown, the CTO, did a UCL PhD on surgical vision and had already worked on FDA and CE marked medical devices. Ayodeji Ijishakin, the COO, built diffusion models for interpretable medical imaging at UCL. Hugo Fry, the CSO, is a Cambridge mathematician and physicist who came out of the MATS interpretability programme.
In October 2024, before the accelerator and before the round, all four appeared as authors on an arXiv preprint called SAE-Rad, with coauthors from UCL and Microsoft Health Futures. It used sparse autoencoders to break a vision model's internals into human readable features and then assembled a report out of them. That preprint is the most independently checkable technical artifact the founders have produced, and it is the root of the interpretability claim the product rests on.
From preprint to booth 5311
Mecha Net v0.1 was announced in February 2025, positioned against finetunes of GPT-4V, Microsoft's MAIRA systems, Google DeepMind's Med-PaLM M and Harvard's MedVersa. Version 0.2 followed in June, trained on fewer than 500,000 chest x-rays where rivals used more than two million.
The company went through Y Combinator's Winter 2025 batch. On November 3, 2025 it announced a $4.1 million seed led by Valia Ventures with participation from Y Combinator, and the same week launched a wider product: multi-anatomy, multi-modality, multi-view, and prior aware, meaning it compares a study against the patient's earlier ones. The roadmap named Mecha Net XR for chest, abdomen, spine, hand, foot, knee, shoulder and hip radiographs, and Mecha Net CT for whole chest CT volumes.
It exhibited at the RSNA annual meeting for the first time that December, booth 5311 in the AI Showcase, on a show floor where more than 200 companies were demonstrating AI.
What is proven, and what is still claimed
| Evidence | What the record shows | Source type |
|---|---|---|
| FDA authorization | No 510(k) and no PMA naming the company as of Sep 24, 2026. No registered trial, no NIH award, no NSF award. | Not found |
| What the FDA has cleared for chest x-rays | Triage and notification software (CADt, product codes QAS and QFM) and computer assisted detection and diagnosis. Lunit INSIGHT CXR Triage cleared as K211733. Aidoc cleared for pneumothorax triage in 2022. Annalise.ai added seven triage findings in 2023. Qure.ai reached 26 cleared indications in February 2026. None of these authorizes a system to produce the radiologist's report. | Public record |
| Regulatory direction | On April 1, 2026 the FDA denied a Harrison.ai citizen petition to exempt four classes of radiology AI from 510(k) review. Most of the 45 plus comments opposed it. | Public record |
| The other side of the Atlantic | Oxipit's ChestLink took CE Class IIb certification in March 2022 to report autonomously on chest x-rays with no abnormalities, clearing the way for deployment in 32 European countries. In the UK, NICE published an early value assessment of chest x-ray AI for suspected lung cancer, and later removed one product, Red dot, from its recommendations because it is no longer available in the UK. | Public record |
| Benchmark results | Self-published. v0.1 reported RadCliQ 2.66 (2.63 to 2.68), ahead of MedVersa at 2.71 and overlapping MAIRA-2 at 2.64. v0.2 reported higher CheXbert micro F1 than all prior models on MIMIC-CXR. No peer review, no external replication. | Company-stated |
| Peer track, founders | SAE-Rad, arXiv 2410.03334, Oct 4, 2024, all four founders with UCL and Microsoft coauthors. | Public record |
| What a funded rival's evidence looks like | MAIRA-X, from Microsoft and Mayo Clinic, was built on 3.1 million studies and tested blind by nine radiologists over 600 studies, reporting critical error rates of 3.0 percent for original reports against 4.6 percent for AI ones. | Independent |
| Named customer | AmeriRad Radiology Solutions, a physician led teleradiology group founded 2023 and licensed in more than 14 states, says it partnered in 2024 to integrate an AI-driven diagnostic overlay. | Partner-stated |
| Unnamed customers | The Y Combinator profile claims partnerships with the largest privately owned radiology practice in the US and a multinational teleradiology company, taking radiologists from one scan an hour to one every five minutes. Neither is named and the only named partner does not describe itself that way. | Claim exceeds the record |
Read plainly: the science is real, the people are credentialed, and the evidence is entirely of a kind the company produced about itself. A radiology report is not a flag on a scan, and in the United States generating one is a device function the FDA reviews. The April 2026 denial says the agency is not loosening that. Europe licensed an autonomous reporting system in 2022 and the UK ran a national evaluation, so the same product can be a cleared device on one side of the ocean and an unapproved one on the other. Mecha Health sells into US practices while that gap is open, with a radiologist signing every report.
What to watch
- Any FDA submission. A 510(k), a de novo or a Q-Submission would be the first sign the company means to be a device rather than a drafting aid.
- Named deployments with published turnaround and error data.
- A Form D or a second round. Nothing has been filed with the SEC under the company's name.
In their words
“Radiologists need tools that don’t ask them to change how they work. They need tools that understand how they work.”
Ahmed Abdulaal, co-founder and CEO, funding release, November 2025 · Company release
“They’re reasoning through anatomy, modality, history, priors, and context all at once. Foundation models are the first technology that can operate at that level of complexity.”
Deji Ijishakin, co-founder, funding release, November 2025 · Company release
“To the best of our knowledge, this is the first work to evidence that radiology foundation models can systematically match or exceed human performance for the task of report generation by analysing medical images.”
Mecha Net v0.2 technical post, June 24, 2025 · Company research
“It presents the first case where a medical diagnostic evaluation will be carried out solely by an artificial intelligence application.”
Gediminas Peksys, CEO of Oxipit, on the ChestLink CE mark, March 2022 · Competitor
“Holding a 510(k) clearance may not reflect that a manufacturer is proficient in, or even has experience with, the processes used in the development of the cleared device.”
FDA decision letter of April 1, 2026, reported by AuntMinnie · Public record
Related companies
Sources
- Public recordFDA 510(k) and PMA queries, no results
- Public recordSEC EDGAR full text search for Mecha Health, zero filings
- Public recordFDA strikes down radiology AI 510(k) exemption arguments
- Public recordFDA denies petition to exempt certain radiology AI devices
- Public recordAI-derived software to analyse chest X-rays for suspected lung cancer
- Public recordAn X-Ray Is Worth 15 Features
- IndependentClosing the Performance Gap Between AI and Radiologists in Chest X-Ray Reporting
- IndependentMecha Health company profile
- IndependentCE Mark for First Autonomous AI Medical Imaging Application
- CompanyMecha Health raises $4.1M seed round
- CompanyCompany and launch pages
- PartnerAmeriRad partners with Mecha Health
Profile researched and written by Healthcare Discovery. Last updated September 29, 2026.
