AI vs. MD: Challenging an MRI Diagnosis with Opus 4.8

After receiving a severe shoulder diagnosis and questionable treatments from a clinic, the author used Opus 4.8 to analyze their raw MRI data. The AI consistently found no evidence of a tear, directly contradicting the human doctor's assessment. This discrepancy has left the author in a state of medical limbo, caught between a human diagnosis they no longer trust and an AI solution that is not yet an official standard.

Key Points

• Human doctors diagnosed a Grade III tendon tear and immediately performed treatments that AI later flagged as clinically questionable.
• The author used Opus 4.8 within Claude Code to process raw DICOM MRI data, a task requiring significant computational work beyond standard chat interfaces.
• The AI's analysis directly contradicted the human diagnosis, finding an intact tendon where the doctor identified a major tear.
• An 'Arbiter' workflow with multiple AI subagents was used to reach a high-confidence conclusion that favored the AI's original finding of no tear.
• The experience highlights a 'trust gap' where AI can expose potential medical over-treatment but remains in an experimental stage for patients.

Sentiment

The overall sentiment is mixed and constructively skeptical. Commenters largely agree with the article that the clinic's behavior deserved scrutiny and that AI can make patients better prepared, but they are much less willing to accept the model's MRI interpretation as medically authoritative. The dominant stance is pragmatic: use AI to ask sharper questions, not to decide whether the radiologist is wrong.

In Agreement

• The clinic's use of a homeopathic injection and aggressive treatment plan makes the author's skepticism reasonable.
• AI can be valuable as a patient-advocacy tool that explains terminology, organizes records, and generates better questions for doctors.
• Doctors and specialists can be wrong, rushed, biased by incentives, or unwilling to revisit earlier conclusions, so patients benefit from independent research.
• Shared decision-making improves when patients arrive informed rather than passively accepting opaque medical authority.
• LLMs may be especially useful for surfacing medical literature, alternative explanations, and inconsistencies that a patient can then verify with professionals.

Opposed

• Generic LLMs are not reliable radiologists and should not be trusted to interpret raw MRI data as a diagnosis.
• Confident, polished LLM output can be inconsistent, leading, or wrong, especially when the user lacks the expertise to verify it.
• Having multiple models or agents agree does not prove correctness because they may share biases, converge on the same error, or follow the prompt framing.
• Medical imaging requires clinical context, calibrated judgment, and accountability that current consumer AI tools do not provide.
• The correct next step is another qualified human medical opinion, not treating AI disagreement as equivalent to professional diagnosis.