A software developer used Opus 4.8 running inside Claude Code to analyze a 266 MB folder of raw MRI DICOM files after receiving a shoulder diagnosis he found suspicious, according to a blog post detailing the experiment. The AI's analysis directly contradicted the clinic's finding of a Grade III partial-thickness tear of the subscapularis tendon.
Processing Raw DICOM Files with Claude Code
After a shoulder injury, the developer's clinic diagnosed a Grade III partial-thickness tear and recommended shockwave therapy and homeopathic injections. Skeptical of the treatment plan, he exported the raw MRI data — a 266 MB folder of hundreds of extensionless DICOM files — and fed it to Opus 4.8 (xhigh) inside Claude Code, giving the model terminal access to install imaging libraries and build its own analysis pipeline.
Given only a minimal prompt describing two to three weeks of shoulder pain, the agent spent roughly an hour analyzing the dataset and produced a 7.72 MB PDF report.
Dissecting the Diagnostic Discrepancy
According to the blog post, Opus 4.8 concluded the patient had mild insertional tendinosis with no discrete tear — the opposite of the clinic's Grade III tear diagnosis. To resolve the conflict, the developer ran an arbitration workflow in Claude Code, using subagents to compare the two findings alongside a prior analysis from GPT 5.5 Pro, which reportedly flagged that the clinic's proposed shockwave treatment violated guidelines for non-calcified tendinopathy. After another hour, the arbiter produced a 4.52 MB report reaffirming, with moderate-to-high confidence, that no partial- or full-thickness tear was present.
Key Data Points
- Data volume: 266 MB DICOM export, hundreds of unstructured file slices
- Agent environment: Opus 4.8 (xhigh) in Claude Code with custom package installation
- Human diagnosis: Grade III (>50%-width) partial-thickness subscapularis tear
- AI arbitration verdict: mild insertional tendinosis, no discrete tear identified
Limitations
The developer, who is not a medical professional, acknowledged that current AI systems cannot guarantee diagnostic precision on complex imaging. The discrepancy leaves him weighing a second human opinion against the AI's more conservative assessment — this is a single self-reported case, not a validated comparison of AI versus clinical diagnostic accuracy.