Medical AI and the Spatial Reasoning Gap: Claude Code vs. the 3D MRI
A viral case study of a developer using Anthropic’s Claude Code (running Claude Opus 4.8) to get a second opinion on an MRI has spotlighted a critical "validation crisis" in medical AI. While the AI’s agentic capabilities allowed it to parse raw medical files and construct a highly detailed, professional-looking report, its final diagnosis flatly contradicted a human radiologist's findings—exposing a dangerous spatial reasoning gap in modern multimodal LLMs.
The Experiment: Parsing 266 MB of Raw DICOM Files
After receiving a diagnosis of a "Grade III partial-thickness tear" in his right shoulder's subscapularis tendon from an orthopedic clinic, developer Antoine decided to seek an AI-driven second opinion. He exported his raw MRI data—a standard DICOM package containing hundreds of files without extensions, totaling 266 MB—and fed them into Claude Code.
Using Claude Code's local execution environment, the agent was able to:
- Install necessary Python libraries (such as
pydicomand image processing tools). - Write custom scripts to parse, extract, and analyze the individual 2D MRI slices.
- Coordinate a multi-agent "arbitration" workflow, where multiple sub-agents analyzed the files independently to avoid bias before an "Arbiter" agent reconciled the findings.
The AI's final, high-confidence verdict was startling: it claimed there was no discrete partial- or full-thickness tear whatsoever, diagnosing only mild insertional tendinosis.
The Spatial Reasoning Gap: Slicing a 3D Medium
The publication of this case study ignited an intense debate among systems engineers, AI researchers, and medical professionals. Radiologists and imaging experts pointed out that MRIs are fundamentally 3D datasets, where the spatial relationship between slices is critical.
Current multimodal LLMs are notoriously bad at 3D spatial reasoning. When an agent writes a script to convert DICOM files into 2D PNGs for analysis, it often slices the data along arbitrary planes, easily missing microscopic features or tears that are only visible when traversing the 3D volume in sequence.
"Especially an MRI which is a 3D medium —something current LLMs are very bad at." — maxall4
"I would not be at all surprised if one could slice an MRI the wrong way to produce a 2D image that fails to show a feature that exists in the source data." — amluto
Furthermore, medical imaging requires microscopic precision that standard vision-language models (VLMs) are not trained or validated to handle. While an LLM can easily identify obvious fractures or large tumors, it struggles with the subtle, pixel-level contrast gradients that distinguish a tear from healthy, albeit inflamed, tissue.
The Human Incentive Crisis
Despite the AI's technical limitations, the case study also highlights why patients are increasingly turning to AI: a deep distrust of human medical incentives.
In this instance, the human clinic had immediately prescribed highly invasive and expensive treatments—including shockwave therapy and Traumeel (a homeopathic medicine "without a therapeutic indication" in Germany)—within minutes of the MRI. A quick query to GPT-5.5 Pro flagged that clinical guidelines explicitly recommend against shockwave therapy for non-calcified rotator-cuff tendinopathy, and that the patient's ultrasound had shown no calcification.
This leaves patients in an uncomfortable "limbo": caught between human doctors who may be over-treating for financial gain, and sophisticated AI agents that write convincing, highly structured, but fundamentally hallucinatory medical reports.