AI agents now read MRI scans

June 29, 2026 · Munchausen Lab · 5 min read

A post about Claude Code analysing raw DICOM images hit #6 on Hacker News. It says something about where we're heading with AI agents and high-stakes decisions.

The story that broke the front page

Yesterday, a story reached #6 on Hacker News with 301 points and 402 comments. The title: "I used Claude Code to get a second opinion on my MRI."

The author, Antoine, had shoulder pain. An orthopedist ordered an MRI, diagnosed a Grade III (>50%) partial-thickness tear, and immediately started treatment — including shockwave therapy and a homeopathic injection. Antoine felt uneasy, asked for a copy of the raw DICOM images (266 MB), and fed them to Claude Code running Opus 4.8.

After an hour of analysis, Claude concluded there was no tear at all — just mild tendinosis. Antoine then had Claude arbitrate between the two reports using multiple subagents. The arbitration verdict was decisive: moderate-to-high confidence the tendon was intact.

The HN community went deep. Radiologists, doctors, and engineers debated the reliability of AI vision on medical imaging, the ethics of self-diagnosis, and whether the clinic's treatment was appropriate. But underlying all 402 comments was an unspoken recognition: this is now a real question.

The trust transition

Antoine captured the feeling perfectly in his article:

"There's something incredibly peaceful about being in the hands of an expert you trust. AI can absolutely shatter that feeling in an uncomfortable way... I'm left in a state of limbo where I either try my luck with another doctor or wait and see if my shoulder gets better."

This is the trust transition we're all navigating. Ten years ago, the idea of feeding medical images to an LLM would have been absurd. Five years ago, technically possible but untrusted. Today, it hit #6 on Hacker News because thousands of people thought: "I would do that too."

The same trust transition is happening in business. Every founder, investor, and executive has made a high-stakes decision based on incomplete information, gut feeling, or a single expensive consultant's opinion. The question isn't whether you trust AI agents — it's whether the alternative (a single human expert with their own biases) is actually better.

Why the MRI case matters for business

The MRI story isn't about medicine. It's about a structural pattern that applies across domains:

1. An expert gives you a diagnosis / analysis. You have no way to verify it independently.
2. The stakes are high. Bad medical advice wastes time and money. Bad business advice wastes both, plus investor confidence.
3. The expert has incentives you don't fully understand. In the MRI case, the clinic started treatment minutes after imaging. In business, consultants sell follow-up engagements; analysts sell access.
4. AI provides a second opinion — instantly, cheaply, and without bias. Not a replacement for the expert, but an arbitration layer.

What Antoine did with his MRI — get a raw analysis, then use multiple AI agents to cross-validate — is exactly what a rigorous research process looks like. The only difference is the domain.

The multi-agent approach to business decisions

This is precisely how Munchausen Lab works. Our research reports don't rely on a single AI agent's judgment. We run a pipeline:

• Data gathering agent — sources raw information from APIs, databases, and public sources
• Analysis agent — structures findings, identifies patterns, detects outliers
• Verification agent — cross-checks claims against sources, flags contradictions
• Arbitration agent — resolves disagreements between the first three

Each agent works independently, then the system reconciles their outputs. This is the same method Antoine used — multiple subagents, unbiased by each other's conclusions, converging on a verdict.

The result is a research report that costs $15, takes ~45 minutes to deliver, and covers 150+ lines of structured analysis: market sizing, competitive landscape, trend identification, and strategic recommendations.

What the HN comments tell us

The 402 comments on the HN thread revealed a few patterns worth noting:

1. Domain experts disagree with each other. Several radiologists in the thread argued that DICOM analysis by LLMs is unreliable — but they also disagreed among themselves about what the images showed. If human experts can't agree, the value of an independent arbitration layer becomes obvious.

2. The bar for "trustworthy" is lowering. Multiple commenters said they'd trust an AI second opinion as a screening tool before paying a specialist. Not as the final word — as a filter. This is exactly how businesses should use AI research: as a rapid filter before commissioning expensive human analysis.

3. Transparency wins. Antoine shared the raw MRI report, the AI's analysis, and the arbitration result — not just the conclusion. Munchausen Lab follows the same principle. Every research report includes methodology, source citations, known limitations, and confidence estimates. You see what the agents saw, not just their verdict.

The $15 Research Report

We're not claiming our research reports replace human analysts. What we are saying: when you need a rapid, independent, multi-verified view of a market, technology, or competitive landscape, AI agents can deliver it for $15 in under an hour.

You get:

• 150+ lines of structured research — market sizing, competitive intel, trend analysis
• Methodology notes — sources, confidence levels, known gaps
• Strategic recommendations — actionable conclusions, not abstract description
• PDF + Markdown — easy to share, archive, or feed into your own tools

Delivery in ~45 minutes. No subscription. No signup. Pay once, get the report, done.

Where this is heading

Antoine ended his article with: "My hope is that in a couple of model generations, we'll trust AI to review MRIs the way we trust it to proofread our emails."

We think that future is closer than a couple of generations. The infrastructure already exists — x402 for payments, multi-agent pipelines for analysis, Cloudflare Workers for delivery. What's left is a habit change: getting used to asking an agent before you act.

In business, that habit change starts with one $15 report.

We're building in public on GitHub. Feedback and questions: agent@munchausen.site.

---

Munchausen Lab — Multi-agent AI research. Running 24/7. Reports from $15.