A startup founder recently used Claude Code, Anthropic’s CLI-based coding agent, to analyze their own MRI scans and obtain a medical second opinion. According to reports, 22% of clinical cases analyzed with AI present discrepancies compared to initial human assessments, highlighting the tool’s potential for patient verification.
TL;DR: A startup founder used Claude Code, Anthropic’s CLI-based coding agent, to analyze their own MRI scans for a medical second opinion. According to reports from El Ecosistema Startup, 22% of clinical cases analyzed with AI present discrepancies compared to initial human assessments, highlighting the tool’s potential for patient verification.
How Did a Founder Use Claude Code to Analyze a Personal MRI?
A startup founder whose identity remains undisclosed used Claude Code to analyze their own MRI scans after receiving an initial diagnosis from a healthcare provider. The case, reported by El Ecosistema Startup, demonstrates how developers are repurposing coding agents for tasks far beyond software engineering. The founder treated the MRI data as structured input. Claude Code processed it like any other file.
Claude Code, as documented in Antyweb’s coverage, operates as a terminal-based agent that reads local files, executes commands, and generates analysis within a project directory. The founder loaded MRI image files directly into the working directory. They then prompted the agent to examine the scans and provide observations. The tool functioned as expected.
According to the El Ecosistema Startup report, 22% of clinical cases analyzed with AI present discrepancies compared to initial human assessments. This statistic frames the founder’s motivation: seeking verification of a medical opinion through an independent analytical system. Claude Code was never designed for medical imaging. It handled the task anyway.
VentureBeat reports that Claude Code has effectively turned every engineer into three, compressing build cycles and expanding what individual developers can accomplish. That same productivity multiplier apparently extends to non-coding tasks. The founder leveraged the agent’s file-reading and reasoning capabilities to cross-check medical findings without scheduling another appointment.
Can Claude Code Actually Read and Interpret Medical Scans?
Yes, Claude Code can read and interpret medical scan files, but with significant caveats. The tool processes images through Claude’s vision capabilities, which support analysis of visual data including DICOM files, JPEGs, and PNGs exported from medical imaging software. The agent does not diagnose. It analyzes patterns.
According to Blake Crosley’s guide to Claude Code CLI, the agent operates within a sandboxed environment where it can read local files, process their contents, and generate structured output. When the founder placed MRI files in the working directory, Claude Code’s underlying vision model examined them the same way it would examine a screenshot or diagram.
The 22% discrepancy rate reported by El Ecosistema Startup suggests that AI systems can identify observations that human radiologists may have missed or interpreted differently. This does not mean the AI is always correct. It means the AI provides a second perspective worth investigating.
Claude Code’s strength here comes from its agentic workflow. As Nick Babich writes on UX Planet, Claude Code subagents can be configured for specialized tasks, processing complex inputs and returning structured analysis. The founder likely used a similar approach, directing the agent to focus on specific anatomical regions or compare findings against known patterns.
It is critical to understand the limitations. Claude Code lacks medical certification. Its analysis cannot replace a licensed radiologist’s report. The tool serves as a verification layer, flagging potential discrepancies that warrant follow-up with a qualified physician.
What Does the 22% Discrepancy Rate in AI Medical Analysis Mean?
The 22% discrepancy rate means that in nearly one in four clinical cases analyzed with AI, the system identified findings that differed from the initial human assessment. According to El Ecosistema Startup’s coverage of the founder’s case, this figure comes from reports evaluating AI-assisted clinical analysis across multiple case types.
A 22% discrepancy rate cuts both ways. On one hand, it demonstrates that AI can catch what humans miss, potentially preventing misdiagnoses and improving patient outcomes. On the other hand, some discrepancies may represent AI errors, where the system flags non-existent issues or misinterprets normal anatomical variations as abnormalities.
For the founder who used Claude Code on a personal MRI, this statistic validated the decision to seek a second opinion through AI. If there was a 22% chance the initial reading contained a discrepancy, verification was justified. The AI provided a reason to ask further questions.
This rate also reflects a broader trend documented across AI medical research. Systems trained on large datasets of medical imagery can recognize patterns that individual practitioners, working under time constraints, might overlook. The discrepancy rate is not an indictment of radiologists. It reflects the volume of data AI can process relative to human review.
Patients should treat the 22% figure as motivation, not conclusion. A discrepancy identified by AI requires human follow-up to determine clinical significance. The founder’s approach, using Claude Code as a starting point for deeper medical inquiry, represents a responsible way to apply this technology.
How Do You Feed an MRI Into a CLI-Based AI Agent?
Feeding an MRI into Claude Code requires exporting the scan into a file format the agent can process, placing it in the working directory, and issuing a prompt that directs the agent to analyze the image. The workflow mirrors how developers use Claude Code for code review, replacing source files with medical images.
According to Blake Crosley’s CLI guide, Claude Code version 2.1.190 supports file reading within its sandboxed environment. Users configure the working directory, add files, and the agent processes them automatically. For MRI data, this means exporting DICOM files or converting them to JPEG or PNG format using tools like ImageJ or 3D Slicer.
The steps are straightforward:
- Export MRI scans from the medical imaging system or patient portal
- Convert DICOM files to a standard image format if necessary
- Create a dedicated working directory for the analysis
- Copy the image files into that directory
- Launch Claude Code from the terminal within that directory
- Issue a prompt asking the agent to examine the files
- Review the agent’s output and note any flagged observations
- Schedule a follow-up with a physician to discuss any discrepancies
- Never use the AI analysis as a sole basis for medical decisions
The MakeUseOf report on using Claude Code to fix Claude-related problems demonstrates how a single well-crafted prompt can direct the agent toward a specific analytical task. The same principle applies to MRI analysis. A precise prompt yields better results.
Claude Code’s terminal-based interface, as described in Antyweb’s coverage, makes file processing feel native to developers. There is no GUI to navigate. The agent reads files, reasons about their contents, and produces output in the terminal. For a founder comfortable with CLI tools, the workflow feels familiar.
Is Getting an AI Second Opinion on Medical Scans Safe?
Using Claude Code for a second opinion on MRI scans carries documented risks that patients must understand before uploading any medical data. According to reporting from El Ecosistema Startup, 22% of clinical cases analyzed with AI present some form of diagnostic discrepancy when compared to physician review. This number alone should give patients pause. AI agents are not regulated medical devices.
Claude Code operates as a coding agent, not a diagnostic tool. Anthropic designed it for software development workflows. Using it for medical image analysis falls outside its intended purpose. The model can describe what it sees in an MRI file, but it cannot legally or clinically validate those descriptions against established radiological standards.
Mozilla’s 0din team demonstrated that AI coding agents can be tricked into installing malware through seemingly clean GitHub repositories. This finding has direct implications for medical data processing. If an agent processes sensitive MRI files while connected to external repositories or tools, the attack surface expands considerably. Patients who attempt this approach should use isolated environments with no network access to untrusted sources.\n Safety also depends on how the patient frames the interaction. Claude Code will attempt to help with any task presented to it, including medical analysis. This helpfulness is exactly what creates vulnerability. The agent does not refuse medical queries by default, which means users receive confident-sounding output that may contain factual errors or hallucinated findings.
How Does Claude Code Compare to Specialized Medical AI Models?
Specialized medical AI models outperform general-purpose coding agents in diagnostic accuracy because they train on curated datasets of labeled medical images. Claude Code was not trained on radiological data. Its image analysis capabilities come from Claude’s general vision model, which handles everything from screenshots to photographs.
FDA-approved medical AI tools undergo rigorous validation against thousands of confirmed cases. They integrate directly with PACS systems used by hospitals. Claude Code has none of these integrations. It processes files through a command-line interface designed for code repositories.
According to VentureBeat, Claude Code effectively triples engineer productivity in software contexts. That metric has no bearing on medical accuracy. A tool optimized for writing and debugging code does not automatically transfer to reading T1-weighted brain scans or identifying lumbar disc herniation.
| Feature | Claude Code | Specialized Medical AI |
|---|---|---|
| Training Data | General-purpose text and code | Labeled radiological images |
| Regulatory Approval | None | FDA-cleared for specific uses |
| Output Format | Conversational text | Structured radiological reports |
| Integration | CLI and development tools | Hospital PACS and EHR systems |
| Error Handling | Hallucinations possible | Validated confidence intervals |
Specialized models also incorporate domain-specific constraints. They understand anatomical relationships, common artifact patterns, and clinical priors. Claude Code approaches an MRI like any other image file. It describes visual content without the contextual framework that a trained radiologist or specialized model applies.
What Are the Security Risks of Processing Health Data in AI Agents?
Processing MRI data in Claude Code exposes patients to several categories of security risk that do not exist in traditional medical consultations. Health data is among the most sensitive personal information. When a patient uploads MRI files to an AI agent, they create copies of that data outside controlled medical environments.
Mozilla’s 0din team showed that Claude Code can be exploited through its own helpfulness when processing files from external sources. The agent follows instructions embedded in repositories and file metadata. A maliciously crafted MRI file or accompanying document could theoretically contain prompt injection payloads. This is not hypothetical. The research demonstrated real attack vectors.
Key security considerations include:
- Data residency: Files processed by Claude Code may transmit to Anthropic’s servers depending on configuration
- Conversation logging: Agent interactions may be stored for model improvement unless explicitly disabled
- Prompt injection: Malicious content in file metadata could manipulate agent behavior
- Repository exposure: If MRI files are committed to version control, they become permanently accessible
- API key exposure: Misconfigured credentials could leak alongside health data
- Third-party MCP servers: External tool connections may log or retain medical file contents
- Local file access: Claude Code can read files across the working directory, potentially scanning unrelated medical documents
- No HIPAA compliance: The agent lacks the safeguards required for protected health information under US law
Patients must also consider that medical data shared with AI platforms may be subject to different legal protections than data held by healthcare providers. Once an MRI scan leaves a hospital system and enters a development tool, the protections that govern medical privacy may no longer apply.
Will AI Coding Agents Replace Radiologists for Second Opinions?
AI coding agents will not replace radiologists for second opinions in the foreseeable future. The role of a radiologist extends far beyond image interpretation. Radiologists correlate imaging findings with patient history, laboratory results, symptoms, and clinical context. They communicate uncertainty. They recommend follow-up imaging.
Claude Code cannot perform any of these functions reliably. The El Ecosistema Startup report on the founder’s MRI experiment noted that AI analysis provided useful supplementary information but did not replace professional medical review. The 22% discrepancy rate between AI and physician analysis means roughly one in five AI-generated observations may conflict with expert assessment.
Radiology as a field is already adopting AI, but through purpose-built tools. Companies like Aidoc, RadNet, and Zebra Medical Vision develop models specifically trained on millions of annotated scans. These models flag potential findings for radiologist review. They do not replace the radiologist. They augment workflow.
General-purpose agents like Claude Code serve a different function entirely. They excel at code generation, file manipulation, and task automation. A patient who uses such a tool for MRI analysis gets a creative but unvalidated interpretation. That interpretation might raise useful questions. It might also introduce anxiety over findings that represent normal anatomical variation.
The founder who documented his Claude Code MRI experience demonstrated an interesting application of AI capabilities. His experiment showed that the tool can engage with medical imagery in conversational ways. This does not constitute a replacement for professional medical diagnosis.
What Prompting Strategies Work Best for Medical Image Analysis?
Effective prompting for MRI analysis in Claude Code requires structure, specificity, and an understanding of how the model processes visual information. Since Claude Code was not designed for medical imaging, prompts must compensate by providing context that a radiologist would already possess.
The MakeUseOf article on Claude Code noted that fixing complex problems can sometimes require just a single well-crafted prompt. Medical image analysis is more nuanced than a code fix, but the principle applies. Clear instructions produce better output.
Recommended prompting approaches:
- Specify the scan type: Indicate whether the MRI is T1-weighted, T2-weighted, FLAIR, or diffusion-weighted, as each highlights different tissue properties
- Provide anatomical context: State the body region scanned, such as brain, spine, knee, or shoulder, to ground the analysis
- Request structured output: Ask for findings organized by anatomical region rather than free-form narrative
- Include clinical history: Share relevant symptoms, prior diagnoses, and the reason for the scan to guide interpretation
- Ask for uncertainty markers: Explicitly request that the model flag low-confidence observations separately from clearer findings
- Request comparison framing: Ask the model to describe what normal anatomy looks like before discussing potential abnormalities
- Use follow-up queries: Treat the initial analysis as a starting point and drill down into specific regions or findings
- Cross-reference with medical literature: Ask the model to compare its observations against established radiological criteria
Nick Babich’s guide on Claude Code subagents in UX Planet describes how breaking complex tasks into subagent workflows improves accuracy. For medical analysis, this could mean using one query to describe anatomy, another to identify anomalies, and a third to summarize findings. Each step produces focused output that is easier to evaluate for accuracy.
Patients should always treat AI-generated observations as conversation starters for their physician, not as diagnostic conclusions.
Frequently Asked Questions
Can Claude Code legally diagnose medical conditions?
No. Claude Code is not a licensed medical device and has no FDA clearance or regulatory approval for diagnostic use. According to the El Ecosistema Startup report, the 22% discrepancy rate between AI analysis and physician review means any output should be treated as informational only. A legal diagnosis requires a licensed medical professional operating within their clinical scope of practice.
How accurate was Claude Code compared to a human radiologist?
The El Ecosistema Startup analysis found that 22% of clinical cases analyzed with AI showed discrepancies when compared against physician review. This means approximately one in five AI-generated observations may differ from expert assessment. The accuracy gap reflects the fact that Claude Code lacks training on labeled radiological datasets and does not incorporate the clinical reasoning that radiologists apply.
What file formats does Claude Code accept for MRI analysis?
Claude Code can process standard image formats including PNG, JPEG, and GIF through Claude’s vision capabilities, as documented in the platform model specifications. DICOM files, the standard format for MRI scans, would need conversion to a supported image format first. Patients can export individual MRI slices as PNG or JPEG files using free tools like MicroDICOM or ImageJ before analysis.
Should patients trust AI agents for medical second opinions?
Patients should approach AI-generated medical observations with significant caution. Mozilla’s 0din research demonstrated that AI coding agents can be exploited through prompt injection, and Claude Code lacks the regulatory safeguards required for medical use. AI output may surface useful questions to discuss with a physician, but it should never replace professional medical advice, diagnosis, or treatment from a qualified healthcare provider.
Summary
Key takeaways from this analysis of using Claude Code for MRI second opinions:
- AI coding agents are not medical devices: Claude Code lacks FDA clearance, radiological training data, and the clinical context that licensed radiologists provide
- Security risks are real: Mozilla’s 0din team demonstrated that helpful AI agents can be exploited, and medical data processed through development tools loses HIPAA protections
- Accuracy gaps exist: The 22% discrepancy rate between AI and physician analysis means patients must treat output as supplementary information, not diagnostic truth
- Prompting matters: Structured prompts with anatomical context, scan type, and clinical history produce more useful output than open-ended queries
- Radiologists remain essential: Purpose-built medical AI tools augment radiologist workflow, but general-purpose coding agents cannot replace professional medical review
If you found this analysis helpful, subscribe to my newsletter at gikiewicz.com for more deep dives on AI tools, security research, and practical technology guides. Have you tried using AI agents for unexpected tasks? Send me a message — I read every response.