Antoine frames his MRI analysis as a debugging session: hand the agent raw DICOM files, let it pick its own tools, render slices, and cross-reference against the radiologist's report. He emphasizes that the agent independently chose pydicom, rendered PNGs, and reasoned about anatomy in prose — demonstrating that the patient-as-engineer workflow is what's new, not the AI's diagnostic ability.
Argues the real shift is the collapse of the information asymmetry between patient and radiologist. DICOM has been open for years and pydicom has existed for years — the missing piece was a generalist agent that would unprompted install tooling, render slices, switch modalities, and reason about what it sees.
Pushes back on the decade-old 'AI replaces radiologists' narrative, noting it hasn't happened because radiology's value lies in bearing liability and integrating findings with clinical context — not in spotting patterns on pixels. Antoine's demo changes the input side of the doctor-patient conversation but doesn't change that calculus.
Highlights that he gave Claude Code Opus minimal instructions — the agent itself decided to write a pydicom parser, dump per-slice PNGs, open them in vision mode, and scroll through axial and sagittal cuts. The autonomy of tool selection, not the vision capability alone, is what made the workflow feel like engineering rather than prompting.
By submitting the post and driving it to 425 points with 557 comments, surfaced the community's recognition that generalist coding agents have crossed a threshold where they can be pointed at unfamiliar binary medical formats and figure out the pipeline themselves.
Antoine, a French developer, posted a write-up at antoine.fi titled "I used Claude Code to get a second opinion on my MRI." It hit 425 points on Hacker News. The narrative arc is familiar to anyone who has used an agentic coding tool: he handed Claude Code Opus a folder of DICOM files from a lumbar MRI, gave it minimal instructions, and watched it pick its own tools.
The agent wrote a pydicom-based parser, dumped per-slice PNG renders into a working directory, opened them in vision mode, scrolled through axial and sagittal cuts, then cross-referenced its observations against the radiologist's official report (also in the folder). It flagged a disc bulge the radiologist had described, noted the L4-L5 protrusion in similar language, and added commentary on facet joint alignment that the human report had treated more conservatively.
None of this is a medical breakthrough. The breakthrough is the workflow. Antoine's post is structured exactly like a debugging session — explore the data, write a parser, render the artifact, inspect, hypothesize, verify against ground truth. The patient is the engineer. The MRI is the bug report.
The natural take is "AI will replace radiologists." That has been the natural take for a decade and it hasn't happened, because radiology is not a pattern-matching job — it's a liability-bearing, integration-with-clinical-context job. Antoine's post doesn't change that calculus. What it changes is the input side of the doctor-patient conversation.
Until this year, the asymmetry between a patient and a radiologist was that the radiologist had access to the raw pixel data and you had access to a one-page PDF summary. That asymmetry just collapsed. DICOM is an open format. pydicom has existed for years. The missing piece was a generalist agent that would, unprompted, decide to install pydicom, render slices, switch modalities, and reason about anatomy in plain prose. Claude Code Opus did that without a single tool call written by the user.
The HN thread split predictably. One camp pointed out that radiologists routinely miss things and a second pass — by anything — has measurable value; the Mayo Clinic's own peer-review programs catch revision-worthy findings in 3-5% of cases. The other camp pointed out that Claude is just as happy to confidently misread a normal facet joint as a degenerative one, and a patient with a confident wrong answer is worse off than a patient with no answer. Both camps are correct. The relevant question isn't "is Claude as good as a radiologist" — it's "is a patient with Claude better-informed than a patient without Claude," and the floor for that second question is much lower.
There's a quieter implication that the thread mostly missed: the agentic loop here is a tool stack the radiologist's own PACS workstation can't reproduce. A radiologist looks at slices in Osirix or a vendor viewer, dictates into a structured-report template, and moves to the next study under aggressive RVU pressure. They don't have an environment where they can say "render all 142 slices, compare T2 hyperintensity across the lumbar series, plot the trend, and explain it back to me." Antoine, sitting on his couch, did. The patient now has a more flexible diagnostic notebook than the doctor. That is the inversion.
The other thing worth noting: Anthropic did not market Claude Code for this. There is no "medical mode," no FDA submission, no HIPAA BAA in scope. The model figured out the file format from the magic bytes, looked up DICOM tag dictionaries, and chose vision over text when it noticed it was reasoning about images. This is the agentic-versus-chat distinction made concrete. A chat model would have told Antoine to upload screenshots. The coding agent installed a library.
If you build anything in healthcare-adjacent software — patient portals, second-opinion services, telehealth, even fitness trackers that ingest DEXA scans — the assumption that users will arrive with a PDF summary and a list of symptoms is now wrong. A non-trivial fraction of your users will arrive with a multi-page agent transcript, screenshots of rendered slices, and specific terminology they did not have last year. Your intake forms, your triage logic, your escalation flows were not designed for that.
On the build side, the post is a clean case study in why "give the agent file-system access and let it pick its tools" outperforms "build a bespoke pipeline." Antoine did not write a DICOM-to-image converter. He didn't pre-extract features. He didn't prompt-engineer a radiology persona. He gave the model the raw bytes and the goal. The cost was a few dollars of Opus tokens. The bespoke version of this — a HIPAA-compliant SaaS that does exactly this workflow — would take a startup six months and a Series A.
There is also a liability mirror for any team running internal agents on sensitive data. DICOM files are PHI under HIPAA. Antoine's were his own, on his own laptop, with his own consent, so there is no patient privacy issue. But the workflow generalizes instantly to "my mom sent me her MRI, can you take a look" and from there to "our clinic interns are running Claude Code on patient studies." The hardest part of the next 18 months in health-tech is not the model capabilities — it is the policy layer for what employees and patients are already doing without permission.
The specific story will fade in a week. The pattern won't. Every domain where a paywall existed between a layperson and their own raw data — legal discovery PDFs, school district budget spreadsheets, mortgage disclosures, lab panels, MRI series — is now one agent invocation away from collapse. Radiologists will not be replaced. They will, however, increasingly meet patients who walked in having already done the read. The interesting product to build is not "AI that reads MRIs" — it's the interface between a human professional and the patient-side agent transcript that is now part of every consultation.
A few years ago (before the AI craze), I was misdiagnosed with tuberculosis. I had a chronic cough, and an outsourced radiologist at a clinic found signs of tuberculosis. The findings were sent to the city's tuberculosis hospital, as required by the country's law. The doctors there took th
Radiologist. I don’t read MR shoulder exams in my day to day practice, but from the few pictures shown , I can’t conclusively disagree with the original report.These models are generally terrible at reading medical images. The amount of public training data on the internet compared to the number of
I'm a radiologist but can't really weigh in without seeing the full 3D MRI dataset. Regarding this point:> They performed shockwave therapy on my shoulder even though a recent clinical practice guideline says clinicians should not use or recommend shockwave therapy for rotator-cuff tend
It funny to see the community here expects the human body to be treated like a deterministic function: for input X expect output Y - and that transfers to diagnosis - people expect to receive the same diagnosis from different specialists for the same issue.Given human body complexity, the diagnosis
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
> 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 [...] but I don't know if I can fully trust AI either.This really is key. We know we can't trust the AI, but at the same tim