A software developer named Antoine has demonstrated the startling capabilities of modern AI by using Anthropic's Claude 3.5 Sonnet to analyze his own MRI scans. The powerful multimodal model successfully identified a potential disc protrusion and nerve compression, offering a "second opinion" that closely aligned with a professional radiologist's official report.
This personal project, detailed in a blog post, highlights a fascinating, if cautionary, use case for publicly available large language models in specialized fields like medicine.
From Raw Scans to AI Insight
To conduct the experiment, Antoine followed a straightforward but clever workflow to prepare his medical data for the AI. He wanted to see if the model could interpret the complex imagery and provide a coherent analysis without any prior medical context.
His process involved several key steps:
- Data Anonymization: He first used software to strip all personal identifying information from his DICOM (Digital Imaging and Communications in Medicine) files.
- Format Conversion: The anonymized DICOM scans were then converted into a standard PNG image format, which Claude can easily process.
- System Prompting: He uploaded the images to Claude 3.5 Sonnet and prompted it to act as an expert radiologist, instructing it to provide a detailed analysis of the lumbar spine images.
Claude's Unsettlingly Accurate Findings
The results were both immediate and shockingly accurate. The AI analyzed the axial and sagittal views of the MRI and pinpointed the specific area of concern. It generated a detailed report that would be difficult for a layperson to distinguish from a human expert's.
Most critically, Claude identified a "moderate-sized left paracentral disc protrusion at L5-S1, causing moderate to severe compression of the transiting left S1 nerve root." According to Antoine, this language and the specific diagnosis mirrored the findings in the official report from his actual radiologist. This demonstrates a powerful capability for AI to not only see but interpret complex, domain-specific visual data.
Navigating the rapid advancements in AI can be challenging, but understanding their real-world impact on industries like healthcare is crucial. Thousands of AI professionals and enthusiasts subscribe to the AI Breaking Wire newsletter for weekly insights that cut through the noise.
What It Matters
Antoine is clear in his post that this experiment should not be interpreted as medical advice and is not a substitute for professional diagnosis. However, this case study is a powerful proof-of-concept for the future of AI as a diagnostic co-pilot. While we are years away from AI replacing radiologists, this shows a near-future where AI tools could empower patients to better understand their own health data and help clinicians triage cases or double-check findings. It's a compelling glimpse into a future where AI makes specialized knowledge more accessible, but it also underscores the urgent need for ethical guardrails and rigorous clinical validation.