Get a second opinion on an MRI
June 28, 2026. Task: The author obtained a DICOM MRI export (266 MB) and clinical treatment records after suspecting premature intervention by an orthopedist. They used GPT 5.5 Pro to flag guideline violations in the treatment plan, then employed Claude Code with Opus 4.8 to analyze the raw MRI files. The AI generated a diagnostic report contradicting the human radiologist's finding of a Grade III tear, reporting an intact tendon instead. A subsequent arbitration step using multiple subagents compared the AI and human reports, concluding with moderate-to-high confidence that no discrete tear existed.; tools: Claude Code. Human role: The original post does not say how a person stayed involved.; limitations: The author explicitly states they are not a doctor and the technology may not be ready for reliable medical diagnosis. The AI's conclusion contradicted the human expert, creating a 'limbo' state rather than a clear answer. Clinical guidelines and treatment appropriateness were flagged by GPT, but the MRI interpretation itself remains unverified by a qualified professional. Users must treat AI medical analysis as experimental and non-authoritative..
Tools
Published on Agent Work Atlas:
Overview
June 28, 2026
Workflow
The author obtained a DICOM MRI export (266 MB) and clinical treatment records after suspecting premature intervention by an orthopedist. They used GPT 5.5 Pro to flag guideline violations in the treatment plan, then employed Claude Code with Opus 4.8 to analyze the raw MRI files. The AI generated a diagnostic report contradicting the human radiologist's finding of a Grade III tear, reporting an intact tendon instead. A subsequent arbitration step using multiple subagents compared the AI and human reports, concluding with moderate-to-high confidence that no discrete tear existed.
Reported result
The AI analysis directly contradicted the human diagnosis, identifying mild tendinosis rather than a severe tear. This discrepancy left the author in a state of uncertainty, unable to fully trust the AI or the initial clinical judgment. The author did not resolve the medical issue via AI but chose to wait and see if rehab improved symptoms, highlighting that AI output currently serves as a source of doubt rather than definitive medical truth.
Human role
The original post does not say how a person stayed involved.
Limitations
The author explicitly states they are not a doctor and the technology may not be ready for reliable medical diagnosis. The AI's conclusion contradicted the human expert, creating a 'limbo' state rather than a clear answer. Clinical guidelines and treatment appropriateness were flagged by GPT, but the MRI interpretation itself remains unverified by a qualified professional. Users must treat AI medical analysis as experimental and non-authoritative.
Try it
I am a patient with [replaceable: specific symptom, e.g., right shoulder pain] who received an MRI and a treatment plan I suspect is aggressive. I have a DICOM export of the MRI and the radiologist's report. Please use Claude Code with Opus 4.8 to analyze the DICOM files. First, install necessary Python packages for medical imaging (e.g., pydicom, nibabel). Then, generate a detailed diagnostic report focusing on [replaceable: specific anatomy, e.g., subscapularis tendon]. Compare your findings with the provided human report: [replaceable: paste human report text]. If discrepancies exist, run an arbitration step using multiple subagents to evaluate the evidence for each claim. Output a final verdict with confidence levels. Do not provide medical advice; this is for informational comparison only. I will review the output before consulting a new specialist.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.