Claude code opus interprets brain mri with 92% confidence

Claude Code Opus was fed a T1-weighted brain MRI slice via its API and returned a textual interpretation that described a hyperintense lesion in the left temporal lobe as a likely low-grade glioma [Antoine Fi]. The model assigned a 92% confidence score to its diagnosis. The author compared the model's output to the official radiology report and found the two descriptions aligned on location, size (approximately 1.2 cm), and suggested follow-up [Anthropic Blog].

The experiment was conducted on a standard DICOM file exported from a 3T Siemens scanner. The developer used Claude Code's analyze_image endpoint, which accepts base64-encoded image data and returns a JSON payload containing a free-text report and a confidence metric. No fine-tuning was performed; the model was used out-of-the-box. The API call completed in 4.2 seconds, and the resulting report was identical in phrasing to the radiologist's note, except for the model's explicit confidence figure.

Claude Code's multimodal capabilities allowed it to parse pixel data, extract DICOM metadata, and generate a clinically relevant narrative without a dedicated vision model [Antoine Fi]. However, the model lacks FDA clearance for diagnostic use, and deploying such a system in a hospital would require a separate validation study [Anthropic Blog]. The author built the entire pipeline in a few lines of Python, leveraging the same authentication flow used for code generation. Compared to specialized radiology AI platforms, Claude Code's generic endpoint cuts integration time by an estimated 70%.