Seeking a Digital Second Opinion: My MRI Experience with Claude Code

By Antoine's Blog | June 28, 2026

This account details my experiment using Opus 4.8 to interpret MRI results and provide a secondary perspective on a medical diagnosis. I recognize that this technology is still evolving, which is precisely why I am documenting this journey—whether it serves as a helpful guide or merely a curious anecdote.

Disclaimer: I am not a medical professional. Please treat the following narrative with a healthy dose of skepticism; this is an exploration of AI, not medical advice.

---

🏥 The Clinical Context

For several weeks, I dealt with persistent pain in my right shoulder. While it seemed to be improving, I sought a professional opinion from an orthopedist. The clinic provided an immediate MRI, which led to the following diagnosis:

Diagnosis: A $\text{Grade III}$ ($\approx 50\% \text{ width}$) partial-thickness tear located at the apical insertion of the subscapularis tendon.

The clinic's proposed treatment plan was aggressive and began almost immediately. However, I left the facility feeling that the medical team had acted too hastily jumped the gun. Before departing, I ensured I had:

1. A full copy of the MRI results.
2. A comprehensive list of all treatments administered and suggested.

Initial Red Flags via GPT 5.5 Pro

Before diving into the images, I ran the treatment list through GPT 5.5 Pro. The AI immediately flagged two concerning points:

• Shockwave Therapy: The AI noted that current clinical guidelines advise against using shockwave therapy for rotator-cuff tendinopathy unless calcification is present. My ultrasound had explicitly shown no calcification.
• Traumeel Injections: I was injected with Traumeel, which is classified in Germany as a homeopathic remedy $\textit{without a therapeutic indication}$.

---

💻 Engineering the Analysis with Claude Code

To analyze the actual imagery, I used a standard DICOM export.

Dataset: MRI_Export_Right_Shoulder
Files: ~Hundreds of extensionless files
Total Size: 266 MB

I opted for Opus 4.8 (xhigh) integrated within Claude Code. For those unfamiliar, the distinction is vital: while Claude.ai is a chat interface, Claude Code allows the model to execute code and install necessary libraries, enabling a much deeper level of technical work.

The Workflow

I instructed the AI to first identify and install any required Python packages for DICOM analysis. I provided minimal context—only that I had "right shoulder pain for 2–3 weeks"—which was actually less information than the human doctors possessed.

Claude's internal checklist:

• Install medical imaging libraries
• Parse DICOM metadata
• Analyze tendon integrity at apical insertion
• Generate comprehensive PDF report

After an hour of processing, the result was shocking. While the doctor saw a $\text{Grade III}$ tear, Opus 4.8 reported that the tendon was entirely intact.

---

⚖️ The Arbitration Process

Faced with two diametrically opposed conclusions, I tasked Claude with an "arbitration" to resolve the conflict. To ensure a fair trial, I provided:

1. The human doctor's report.
2. The initial AI analysis.
3. A transcript of my conversation with GPT 5.5 Pro regarding physical movements and positions to test the injury.

The Methodology

Claude employed a sophisticated approach using multiple subagents to prevent "context bias," ensuring each analysis was independent before being synthesized.

The Final Verdict

The resulting Right-Shoulder-MRI-Arbitration.pdf was decisive:

Arbiter's verdict: Evidence favours Reader A (moderate-to-high confidence).

Finding: Mild insertional tendinosis; NO discrete partial- or full-thickness tear identified, including at the apical insertion.

---

📉 Comparison Summary

---

💭 Final Reflections

It is fascinating—and unsettling—that two "experts" could look at the same data and reach such opposite conclusions. There is a profound sense of peace when you trust a human expert implicitly; you simply follow their lead. AI has the power to shatter that peace.

Now, I am stuck in a state of medical limbo:

• Do I seek a third opinion from another human doctor?
• Do I trust the AI and simply wait for my current rehab to work?

I hope that in a few iterations, AI medical reviews will be as trusted and routine as using AI to proofread an email. Until then, I remain a curious observer of my own health.

Note: I have omitted the names of the clinic and physician, as this is a study of technical curiosity, not a complaint.

#ai #experiment #health