Substrate
Medical AI
An agent layer inside the OHIF medical imaging viewer. The radiologist reads the images. The agent arranges studies, navigates, applies display presets, and drafts findings that cite accepted measurements. Research use only.
Key features
- Ten WebMCP tools, four that read the viewer's state and six that change it
- No tool returns pixel data or accepts coordinates chosen by the agent
- The agent can hang two timepoints side by side, move to a slice, and apply window and level presets
- A proposed measurement must copy one the radiologist already drew
- Only the radiologist can accept a proposal, sign, or export
- Every change the agent makes can be undone
- Research use only. The FDA has not cleared it
Substrate adds an agent to the stock OHIF viewer. The split is simple. The radiologist does the diagnostic work. The agent does the work around it: finding the prior study, putting the two side by side, setting lung windows, going to slice 80.
Why not browser automation
An imaging viewer holds state an agent can’t reliably read off the screen: which studies are open, which slice each viewport is on, the display settings, the measurements. Substrate hands that state to the agent through ten named tools, so the agent never has to guess from pixels or click by coordinates.
What the agent can’t do
It can’t interpret an image, invent a finding, accept its own proposal, sign, or export. A draft report can only cite measurements the radiologist has accepted.
Try it
There is a live demo on a public chest CT study, a video, and the code. It is for research use only and must not be used for diagnosis.