Mecha Health (YC W25)

Can we peek inside a radiological foundation model—and actually understand what it’s thinking? Turns out… we can. And many of the features the model learns are surprisingly human-readable. In our latest work, we used sparse coding techniques (inspired by Anthropic’s research on monosemanticity) to reverse engineer the internal representations of a radiology foundation model. The model had learned clinically relevant, interpretable features—like surgical hardware, cardiac assist devices, and the concept of clear lung fields. We further demonstrated that a foundation model’s ability to learn clinical features is not uniform, and is likely impacted heavily by data quality. These are promising results which suggest that interpretability techniques could be used to better understand how AI systems make their decisions, and to ensure they’re reliable. We’re also proud to see our earlier work cited in Anthropic’s recent paper on “tracing the thoughts” of language models. We look forward to continuing to push the boundaries of interpretable AI in medical imaging. Excited about the future of interpretable AI in radiology? Let's connect! Link to the article in the description. #Radiology #MedicalImaging #AIinHealthcare #ExplainableAI #FoundationModels

We had the privilege of hearing from one of our Patron Saints—Paul Graham, at the original YC building on Pioneer Way. A key takeaway? Delight your users by delivering overwhelming value. At Mecha Health, that means one thing: exceptional clinical accuracy. But how do you ensure it? First, rigorous evaluation—measuring output quality to optimize toward a well-calibrated target. As we discussed in our first research post, even knowing what a good radiology report looks like is difficult to define. We are actively developing evaluation metrics and are excited to share these with the community. Second, new algorithms—current approaches to draft report generation fall short, and even the best systems today lag behind human accuracy. Despite the massive compute budgets of big players, we are still in the GPT-2 days of report generation, so fundamental technical breakthroughs are required. That's why we're relentlessly committed to research and engineering. We are excited to create powerful, verifiable systems and to redefine how they’re evaluated. If you're a health exec who’s eager to join us on our mission to build the next generation of radiology AI, let’s talk! #RadiologyAI #Healthcare #MedTech #MachineLearning #AI #Innovation #WinningwithMechInterp

They say never meet your heroes—but what if meeting them changes everything? The Mecha team bumped into Yann LeCun on his way to receive the Queen Elizabeth Prize for Engineering alongside Yoshua Bengio, Geoffrey Hinton, John Hopfield (Foundations), Bill Dally, Jensen Huang (Hardware), and Fei-Fei Li (Datasets). His work in computer vision and deep learning stands alone and we build using his innovations every day. Even though our meeting was brief, he was very encouraging about the problems we’re solving. At Mecha Health, we want to ensure instant access to world-leading radiological expertise. It is clearly unacceptable to wait days for diagnostic or management decisions. It is clearly unacceptable for health systems to cause avoidable patient harm. There are many scientific and technological challenges ahead to create the sorts of AI systems required to mimic and assist human reasoning in healthcare. Yann doesn’t believe current systems are even close. But we’re very excited to be building in this space. If you’re a tech-forward clinician or healthcare executive - join us in tackling these challenges. Haters will say they can’t spot our final co-founder, Hugo Fry. #AI #DeepLearning #MachineLearning #HealthcareAI #Radiology #Innovation #Leadership #FutureOfHealth

🚀 Introducing Mecha Net v0.1: The Best Chest X-Ray Reporting Model in the World. 📡 Most AI models for Chest X-ray (CXR) interpretation today work on a narrow scope—detecting just one or two abnormalities at a time. 💡 That’s where Mecha Net comes in. Instead of focusing on isolated findings, it’s a generalist system designed to generate full radiology reports from a scan. This is a huge leap forward in automating medical imaging analysis. 📊 How does it perform? We compared Mecha Net to the current state-of-the-art (SOTA) models in CXR report generation. In particular, we compared to finetunes of OpenAI’s GPT-4V model, the MAIRA systems from Microsoft, Deepmind’s Med-PaLM M, and Harvard’s Medversa. The results are: ✅ Best-in-class performance on CheXbert F1 scores across all 14 clinically significant findings. ✅ Outperforms prior models—even in early versions without auxiliary data (no patient history, prior scans, segmentation masks, lateral views, or patient demographics). ✅ With auxiliary data (just the scan’s indication + up to two prior text reports), Mecha Net 0.1 sets the new SOTA for this benchmark. 🔬 What’s next? We’re working towards a future where no one has to wait for their scan results ever again. That means: 🩺 More clinical trials with direct radiologist feedback 🛠️ Pushing the underlying models even further 📈 Optimizing how auxiliary data improves accuracy We’re just at version 0.1—we have a lot of work to do to get to 1.0, but we’re completely driven to find the limits of deep learning for medical imaging. We’re at a stage of setting up pilots and first integrations. We are selecting a limited number of clinical/design partners. If you are at an AI-forward radiology practice, tele-radiology company or health institute - stop considering whether you should reach out, and reach out right now! We’d love to hear your thoughts in the comments! Link to article below. #AIinHealthcare #MedicalImaging #Radiology #ChestXray #DeepLearning #Transformers #FoundationModels #HealthcareInnovation

We are looking forward to attending The Royal College of Radiologists and NHS Global AI Summit this Tuesday 4th February in our native London, UK. We have been working hard to push our science and are excited to share the latest breakthroughs at Mecha Health. We also look forward to learning from fellow innovators in the field and are particularly keen to connect with radiologists interested in exploring the future of clinical workflow automation. If you spot Ahmed or Ayodeji at the summit, please say hi! We'd love to exchange insights on where radiology AI is heading in 2025 and explore potential collaborations. Also feel free to reach out via LinkedIn! #HealthcareAI #Radiology #MedicalImaging #ArtificialIntelligence #HealthTech #NHS #Innovation

🤔 AI-generated radiology reports are nearly here—but how do we know they’re clinically reliable? 🩻 At Mecha Health, we care a lot about this problem, and have been thinking hard about what the current evaluations mean, and if they’re even useful in the first place. This week, we released our first research blog, where we explore the current metrics that evaluate AI-generated reports and share our vision for the future of the field. 📑 Automated report generation has the potential to transform radiology by reducing workload and improving patient care. Ensuring that these reports are hallucination-free and clinically accurate is crucial to ensure this potential is positively realised. As a result, several radiology-specific automatic evaluation metrics have been proposed such as CheXbert, RadGraph-F1, RadCliQ and FineRadScore. But evaluating free-text reports is incredibly difficult due to the complexity and context-dependency of radiology. Some of the fundamental problems include: ⚠️ Hallucinated detail can appear correct but still lack support. ⚠️ There is a high variability in how radiologists write reports. ⚠️ Ground truth reports might omit or otherwise miss correct findings. We explore these challenges and propose the Clinically Aligned Radiology Evaluation (CARE) score. The CARE score allows for: ✅ A decomposition of reports into clinical statements, and a shared ontology between semantically equivalent statements. ✅ Bi-directional entailment between ground-truth and generated reports, to assess whether statements have been hallucinated or omitted. ✅ Awareness of the clinical severity of errors in generated reports. 🤓 We’re currently developing this score further for real-world clinical impact, and welcome collaboration from relevant researchers. 🔗 Check the link the comments for our research post. And let us know your thoughts on the future of evaluating AI-generated radiology reports. This work is based on the amazing talks given by Pranav Rajpurkar, Matthew Davenport MD, and Woojin Kim at RSNA 2024 last month. #AI #Radiology #MachineLearning #Healthcare #HealthcareInnovation

Mecha Health (YC W25) produces trustworthy medical image reports using AI, improving reporting accuracy and efficiency for healthcare professionals. They're creating instant, global access to the world’s best radiological expertise— no one has to wait for a scan ever again. Due to a shortage of experts, US radiologists have to read 100 - 250+ studies per day, and this means that 55% have reported high levels of burnout. Additionally, patients sometimes have to wait days to receive medical reports and diagnoses from their scans due to backlogs. At times, this can result in preventable patient harm. Mecha Health has invented a new way to generate draft medical reports in a way that is accurate and interpretable. Their system can trace every finding to its visual source in the image. This allows for transparent reporting that doctors can trust. Ahmed Abdulaal, a trained medical doctor from Imperial College London, saw firsthand how reporting delays harm patient outcomes. He met co-founders Nina Montaña Brown, Ayodeji Ijishakin, and Hugo Fry through University College London’s Medical Imaging PhD program. As leading researchers in machine learning and medical imaging, having published in tier-one AI venues (NeurIPS, ICLR, ICML), they're on a mission to transform radiology. Congrats to the team on the launch! 🚀 https://lnkd.in/gN9tZ9kS

🎉 We're thrilled to announce the launch of Mecha Health, backed by Y Combinator! Mecha Health is creating instant, global access to the world’s best radiological expertise. No one waits for a scan ever again. 🏥 Due to a shortage of experts, US radiologists have to read 100-250+ studies per day, with 55% reporting high levels of burnout. Patients sometimes wait days to receive medical reports and diagnoses from their scans due to backlogs in radiology. These delays can lead to preventable patient harm. 💡 Our innovation: We've invented a new way to generate draft medical reports that is both accurate and interpretable: ✅ Our AI system traces every finding directly to its visual source in the image. ✅ Each finding is evidenced with examples from the training dataset ✅ This enables transparent and accurate reporting that doctors can trust 👥 Founded by Ahmed Abdulaal, Nina Montaña Brown, Ayodeji Ijishakin and Hugo Fry – leading researchers in machine learning who crossed paths through University College London's Medical Imaging PhD program. Ahmed, a trained medical doctor from Imperial College London, witnessed firsthand how reporting delays harm patient outcomes—driving Mecha Health's mission to transform radiology. 🤝 We're looking to connect with: Clinicians and radiologists Healthcare executives Radiology practices and imaging centers Help us transform radiology and improve patient outcomes. 📧 Reach out: ahmed@mecha-health.ai 🌐 Learn more: mecha-health.ai #HealthTech #AI #Radiology #YCombinator #HealthcareInnovation #MedTech