Using Claude Code and Opus 4.8 for MRI Analysis: A Technical Case Study

Executive Summary

Using Claude Code with the Opus 4.8 model to analyze MRI data can produce highly confident but potentially inaccurate results that contradict professional medical diagnoses. While AI can synthesize complex medical reports and offer a structured approach to data analysis, it currently lacks the specialized training and spatial reasoning required for reliable radiological interpretation.

The Technical Workflow: Analyzing DICOM Data with Claude Code

Analyzing medical imaging requires more than a standard chat interface. In this case study, the user utilized Claude Code, a tool that allows the model to execute code and install packages, rather than the standard Claude.ai chat. This distinction is critical because it enables the model to programmatically handle large datasets—such as a 266 MB DICOM export containing hundreds of extensionless files—which would be impossible to upload to a standard chat window.

Implementation Strategy

• Environment Setup: The model was instructed to install necessary Python packages for DICOM processing and image analysis.
• Iterative Planning: The user provided minimal context ("right shoulder pain for 2–3 weeks") and required the model to develop a detailed execution plan before taking action.
• Multi-Agent Arbitration: To resolve contradictions between the human doctor's report and the AI's initial finding, the user implemented an arbitration workflow. This involved using multiple sub-agents to perform independent analyses to avoid bias from existing context.

Conflicting Results: Human vs. AI

The primary objective was to verify a diagnosis of a "Grade III (>50%-width) partial-thickness tear at the apical insertion" of the subscapularis tendon.

The AI Verdict

After an hour of processing, Opus 4.8 reported that the tendon was intact, directly contradicting the human orthopedist. In a subsequent arbitration phase, the AI remained decisive, concluding that evidence favored its own initial finding of "mild insertional tendinosis" with no discrete tear identified.

The Human Verdict

The clinic's orthopedist diagnosed a significant partial-thickness tear and recommended an extensive treatment plan, including shockwave therapy and homeopathic injections (Traumeel).

Expert Critique and Technical Limitations

Professional radiologists and technical experts in the community have highlighted several reasons why LLMs currently fail at this task:

1. Lack of Specialized Training Data

Radiologists emphasize that the volume of public image-text pairs (scans paired with professional reports) available for LLM training is minuscule compared to the thousands of scans a human radiologist reads during residency.

2. Spatial Reasoning and Tokenization

LLMs do not "see" images as humans do. Images are tokenized, which often leads to failures in spatial recognition and precise localization of anatomical anomalies. One expert noted:

"Unless you can turn the image into a natively tokenized format like JSON or something that somehow accurately tokenizes what's on there, I would NOT trust Dr. Claude's analysis."

3. The "Annotation Trap"

Some users reported that LLMs may not be analyzing the image pixels at all, but rather reading the embedded metadata or annotations within the DICOM container format, leading to a false sense of competence.

Synthesis of Insights: The Role of AI in Healthcare

While the AI failed the radiological test, the discussion revealed several areas where LLMs provide genuine value in a medical context:

• Report Translation: Converting dense medical jargon into human-readable summaries for patients.
• Brainstorming and Research: Surfacing niche studies (e.g., NIH studies on localized allergic reactions) that a primary care physician might overlook.
• Patient Advocacy: Helping patients identify when a treatment (like shockwave therapy for non-calcified tendinopathy) contradicts established clinical practice guidelines.

Conclusion

AI-driven medical analysis currently exists in a state of "competence vs. accessibility." While LLMs are more accessible and sympathetic than time-constrained doctors, their lack of liability and potential for hallucination make them dangerous for primary diagnosis. The consensus among experts is that AI should be used as a brainstorming tool for the patient and the doctor, but never as a replacement for a certified radiologist.