Methodology
How the AI reads a scan — and how accurate it is
Medical AI is only useful when it shows its work. This page names the models, walks through the pipeline, publishes the accuracy caveat most products hide, and lists the limits we know about. Your radiologist’s report stays authoritative throughout; Read Your Scan is informational, never a diagnosis.
1 · De-identify in your browser
Identifying DICOM tags — name, IDs, dates — are stripped on your device before anything is uploaded. The models never see who you are.
2 · Read every series
The exam is grouped into its acquisitions and each one is read slice by slice by a frontier vision model, with the anatomical plane derived from the volume’s own geometry.
3 · Merge, grade, synthesize
Findings seen on several series are merged; a separate pass grades severity from the reading’s own wording; a final pass writes the plain-language report. Every finding keeps a confidence level and a clickable location on your images.
4 · Cross-check (Second Opinion)
On request, a second independent model reads the same exam and the two readings are shown side by side — agreement and disagreement both visible.
The accuracy number, stated plainly
Today’s general medical vision models score roughly 60% on published imaging benchmarks. Hiding that behind a polished report would be the easy thing to do. Instead, every finding carries a confidence level, the Second Opinion shows two models’ readings side by side, and your radiologist’s written report — when you upload it — is treated as the ground truth the AI helps you understand, never something it overrules.
Methodology questions, answered
- Which AI models read the scan?
- Deep Analysis is read by frontier vision models — currently Anthropic’s Claude, with Google’s Gemini as an independent reader — over the full de-identified volume, series by series and slice by slice. A separate pass turns the structured findings into the plain-language report. The exact models used are recorded with every analysis.
- How accurate is an AI reading of medical images?
- Honestly: moderately. Today’s general medical vision models score roughly 60% on published imaging benchmarks. That is useful for understanding and orientation, and nowhere near a radiologist. This is why the product treats your radiologist’s written report as the ground truth when you upload it, and why the Second Opinion runs more than one model instead of presenting a single confident answer.
- Why two independent readers?
- Two models that independently land on the same finding are a much stronger signal than one model sounding sure. Where they disagree, the disagreement is shown — a finding only one reader saw is exactly the thing to raise with your doctor.
- What does the AI actually receive?
- Rendered slices of your de-identified volume — identifying DICOM tags are stripped in your browser before anything is uploaded. The reading covers every diagnostic series of the exam, not a single screenshot.
- How is the severity score computed?
- When your radiologist’s report is mapped, severity echoes the report’s own wording — it never invents a new clinical judgment. For AI readings, a separate text-only pass grades the reading’s own sentences, and the score is labelled as coming from the AI, never mixed with the report’s.
- What are the known limits?
- An AI reading can miss findings, under-call severity, and mis-rank which level of the spine is most affected. It reads downsampled slices, so small structures can escape it. It is informational only — not a medical diagnosis, not a medical device, and not a substitute for a licensed radiologist or your doctor.
- How do you test changes to the reader?
- Against ground truth. We keep an evaluation harness that runs candidate readers and prompts on real, consented, de-identified exams that also have a radiologist’s written report, and scores the output against that report — which findings and levels it named, how specific the language is, and whether anything anatomically impossible slipped in. Reader changes ship only when they measure better.
See it on your own scan
Drop your scan and your radiologist’s report on the homepage — the report is mapped and explained free. The independent AI reading is a one-off $9 per exam.