A Developer Used Claude Code to Get a Second Opinion on His MRI

A developer who writes on his personal blog antoine.fi has published an account of using Anthropic's Opus 4.8 inside Claude Code to analyze the DICOM files of his own shoulder MRI and produce what amounts to a second opinion on the diagnosis he received at a clinic. He stresses that he is not a doctor and that the exercise was driven by technical curiosity.

How the experiment started

The story begins with several weeks of pain in his right shoulder. An orthopedist recommended an MRI, which the clinic had on site. The report described a „Grade III (>50%-width) partial-thickness tear at the apical insertion“ of the subscapularis tendon; the clinic started treatment minutes after the scan and suggested repeating it three times in total. He asked for a copy of the results and a list of the treatments performed.

He first sent the paperwork to GPT-5.5 Pro, which immediately flagged two things: shockwave therapy had been applied even though a recent clinical practice guideline advises clinicians not to use or recommend it for rotator-cuff tendinopathy without calcification — and the ultrasound showed no calcification — and he had been injected with Traumeel, a product registered in Germany as a homeopathic medicine „without a therapeutic indication“.

Running Opus 4.8 on 266 MB of DICOM

The MRI package was a standard DICOM export: a few hundred files with no extensions, around 266 MB in total. To analyze it he ran Opus 4.8 at high reasoning effort inside Claude Code rather than in a chat window, so the model could execute code and install the packages required. His only prompt was „right shoulder pain for 2–3 weeks“ — less context, he noted later, than the doctors had.

Roughly an hour later the model produced a 7.72 MB PDF report. Its central finding contradicted the clinic: where the doctor had seen a Grade III partial-thickness tear, Opus 4.8 reported an intact tendon. He called that disconcerting, having expected a lower grade.

Arbitrating the two readings

To resolve the conflict he asked Claude to compare the two reports, adding a separate conversation with ChatGPT 5.5 Pro about movements and positions that probed the diagnosis. The plan was methodical and used multiple subagents to keep fresh analyses free of the existing context. After about another hour it returned a 4.52 MB arbitration PDF: „Evidence favours Reader A (moderate-to-high confidence). Mild insertional tendinosis; NO discrete partial- or full-thickness tear identified, including at the apical insertion.“ The author notes the arbitration named the disputes it could not settle and still ruled decisively here.

Where that leaves the patient

The episode left him, as he puts it, in a state of limbo. The diagnosis and treatment plan look premature and more intervention-heavy than the evidence justifies, but he is not ready to fully trust the model either. He continues physical therapy and may seek another doctor if the shoulder stops improving, is not naming the clinic or physician, and repeats that none of this is medical advice. His hope is that in a couple of model generations people will trust AI to review MRIs the way they trust it to proofread email.