Clinical AI: faster triage from medical imaging
Early-stage AI workflows to speed the triage of head-trauma cases from CT and other diagnostics, surfacing the time-critical findings first — built alongside clinicians at a leading public hospital.
The mandate. A leading public hospital wanted to speed the triage of head-trauma cases, so the most urgent — a large bleed, a midline shift, rising pressure on the brain — surface sooner.
Our role. We built early-stage AI workflows alongside the clinicians who would use them, ordered around the findings that actually change urgency.
The principle. Triage, not diagnosis. A human stayed at every decision, and their corrections fed back into the system.
What the work involved
The goal was triage, not diagnosis — workflows that order a queue of head-trauma cases so clinicians reach the most time-critical first, not a system that tries to make the call for them.
In a trauma pathway the first checks happen at the bedside, but the decisions that cannot wait usually turn on what the CT shows: a large haemorrhage, a shift in the brain’s midline, the secondary signs of rising pressure. Those are the findings that move a case to the front of the line, and surfacing them sooner was the whole point.
We built it early-stage and in the open with the clinical team, so the people responsible for the call stayed in the loop on how the workflow reached its ordering.
Triage, not diagnosis — the people responsible for the call stayed in the loop on how the queue was ordered.
What the workflow looked for
We worked with clinicians to pin down the findings that genuinely change urgency, rather than trying to read everything at once. On a head CT that meant the signals a radiologist would prioritise themselves: intracranial haemorrhage — where it sits, how large, whether its borders read as acute or chronic; midline shift, measured in millimetres against the brain’s own landmarks; the secondary signs of raised pressure, such as compressed ventricles, effaced cisterns, sulcal effacement and herniation; and skull fractures, with their type and displacement.
Encoding the clinicians’ own sense of what matters kept the ordering legible to the people relying on it. The queue moved for reasons a radiologist could recognise and check, not for reasons hidden inside a model.
How it ran
We worked alongside clinicians at a leading public hospital, shaping the workflows around how triage actually happens rather than around the model.
A CT study passed through an inference step that segmented and annotated the scan; a neuroradiologist then reviewed what came back. Where the read held, it fed a draft the clinician could act on; where it did not, their own annotations went back into improving the system. The loop was the point — every case kept a human at the decision, and each correction made the next read better.
Deployment and privacy
The build was designed to live inside the hospital’s own environment. Scans were anonymised before anything left the premises, analysis returned as annotated images into the systems clinicians already use, and the whole workflow could run on infrastructure the hospital controlled.
The same approach supported two modes: a complete read for settings without radiology cover, or a faster second pair of eyes where a team already has one — computer-aided triage that makes an existing workflow quicker without taking the call out of clinical hands.
