Every read leaves a trace.
InferenceMD records how a radiologist moves through a study and what they dictate, then links the two: which slices held their attention, and which findings they spoke about those slices.
Ventricles are normal in size and configuration.
No extra-axial collection over the convexities.
Posterior fossa structures are unremarkable.
Partial opacification of the left mastoid air cells.
Labeled slices from the reads radiologists already do.
Replay how any case was read.
InferenceMD captures how a radiologist scrolls through a study and what they dictate, and turns it into slice-level labels for imaging AI, case replays for teaching, and reading analytics for research.
InferenceMD captures how a radiologist scrolls through a study and what they dictate, then replays the read slice by slice with the dictation in step. The same capture gives slice-level labels for imaging AI and reading analytics for research.
Also from the same capture
Labeled slice map
Each study comes back with the slices that held attention and the dictated findings pinned to them: weak labels for imaging AI without a separate annotation pass.
Small chronic lacunar infarct in the left basal ganglia.
Attention by slice
Findings linked to slices
Show and Tell case replay
Replay a read slice by slice with the dictation alongside, for tumor boards, teaching and case conferences.
Reading analytics
Compare how different readers covered the same study, find past reads that resemble this one, and study review patterns such as automation bias.
Review coverage
Time spent on each slice, two readers, same study
Similar reads
Past reads whose navigation and dictation most resemble this one
Training signal from reads that already happen
Attended slices linked to dictated findings give imaging AI a label source that needs far less hand annotation.
No annotation step for the radiologist
No outlining, no clicking through label menus. The signal comes from scrolling and dictation.
No patient identifiers needed
The method runs on the navigation trace and the dictation alone.
How it works
The labels come from the read itself, from scrolling and dictation that already happen, which is why they cost far less than manual annotation.
Capture
Record which slice is on screen, moment by moment, while the study is read, along with the dictation.
Link
Match each dictated finding to the slices the radiologist was looking at when it was spoken.
Use
Export an enriched record per study: the trace, the slices that held attention, and the findings tied to them.
What teams use it for
- Training data for imaging AI
- Slice-level weak labels drawn from real reads, for models that would otherwise wait on manual annotation.
- Show and Tell
- Replay how a case was read, slice by slice with the dictation alongside, for teaching and case conferences.
- Reading-pattern research
- Similarity search across reads, adequacy-of-review measures, and studies of automation bias.
See a read captured
A 30-second walkthrough of the prototype, built as a 3D Slicer extension.
Demo video coming soon. Book a call to see it live.
Built for teams working with imaging data
- AI-first PACS vendors
- Data and labeling companies
- Academic imaging AI labs
- OHIF and viewer developers
- Radiology groups
Meeting at RSNA 2026?
We'll be in Chicago, November 29 to December 2. Book 30 minutes to talk through your data, your viewer, or your research question.
Book a 30-minute callInferenceMD was founded by David Vu, MD, a neuroradiologist, and Mike McGovern, an ML engineer. Questions: support@inferencemd.ai