Healthcare

The scan that cannot wait, read first.

After a crash, the CT of the head, neck, chest, and abdomen is the moment where minutes decide outcomes. In the regions where crashes are deadliest, the specialist to read it is often hours away. We evolve a triage model that puts the bleeding brain at the top of the queue, on the scanner where it was taken.

The problem

Road deaths cluster where radiologists are scarcest.

The long highways between cities produce the worst crashes and deliver them to the smallest hospitals.

2xthe national road death rate in the worst affected regions of one country, with the lowest radiologist density in the same regions
4 in 10serious crashes happen outside cities, far from trauma centers and overnight specialist cover
Minutesseparate a survivable brain bleed from a fatal one. The scan is fast. The read is the bottleneck.

A trauma CT is a large study: hundreds of images across the head, cervical spine, chest, abdomen, and pelvis. At a trauma center a radiologist reads it within minutes and a surgeon acts on it. At a regional hospital at two in the morning, it is sent to a teleradiology hub and waits in a queue behind every other scan that arrived that night. The patient with a slowly expanding bleed looks stable until they are not.

Triage AI for trauma exists, but it was built for the trauma center. It assumes modern multi detector scanners, standard protocols, and a radiologist close by to confirm. The regional hospital that needs triage most has an older scanner, a different protocol, and a reader who will see the study in an hour if it is urgent and in six if it is not. Nobody told the model which hospital it is in.

In plain terms

Imagine an emergency room where every patient takes a number and waits in order of arrival. The person with a headache and the person with a bleed inside their skull hold the same number. Triage is a nurse who looks at everyone in the waiting room and pulls the bleed to the front. We do that for the scans.

Why the standard approach fails

Triage models are tuned for the hospitals that need them least.

01

Older scanners, thicker slices

A small bleed that is obvious on a modern scanner with thin slices can be a faint smudge on an older machine. Models trained on the former miss it on the latter.

What we do. The model is evolved on studies from the actual scanners in the region, and judged by its performance on the oldest one.

02

A missed bleed and a false alarm are not equal

Standard models balance the two. In trauma, a missed bleed can be fatal and a false alarm costs a radiologist ten minutes. The balance should not be even.

What we do. Recall on time critical findings is a hard floor. The false alarm rate is a tradeoff the trauma team sets, per hospital, based on how much reading capacity they have.

03

Trauma is many findings, not one

Brain bleed, skull fracture, cervical spine injury, collapsed lung, ruptured spleen, pelvic fracture. A product that flags one of them leaves the rest in the queue.

What we do. The whole study is read for the findings that change the next hour, and the study is ranked by the most urgent one in it.

04

The queue is the product

Flagging a scan is not the same as getting it read sooner. Most tools stop at the flag. The hospital still has one reader and forty scans.

What we do. The output is a reordered worklist with a time to read target per study, and an alert path to the on call team when the queue cannot meet it.

A miss and a false alarm do not weigh the same. Standard models are tuned as if they do. In trauma, a missed bleed can cost a life and a false alarm costs a radiologist ten minutes. We set the floor on the miss and let the trauma team decide how many false alarms they can afford.
Missed bleed False alarm Can cost a life. Floor set here. Costs ten minutes. Budget set by your team.
How we solve it

Evolved for the scanner at the end of the highway.

The same evolutionary loop as our other imaging systems, with the objectives trauma demands and the hospitals trauma reaches.

In plain terms

We start with models that already read head and body CT. We make dozens of versions and test them on crash scans from the regional hospitals themselves, including the old scanners. Versions that miss a bleed are thrown out immediately. Among the rest, the ones that raise the fewest false alarms and read the fastest become parents of the next generation. After a few rounds, the model is tuned to the scanner at the end of the highway, not to a trauma center in the capital.

Not one model. A population that is bred. Every generation is made from the last: the strongest candidates become parents, their settings are combined, small variations are introduced, and the children are scored on your own cases. Weak lines die out. Strong ones carry forward.
Generation 1Generation 6Generation 14 48 candidates, all differentWeak ones gone, strong ones combinedYour model, chosen by your team
This is an evolutionary algorithm, not a leaderboard. We do not train many models and keep the best. Each generation is bred from the last. The strongest candidates are selected as parents, their settings are combined to make children, small random changes are introduced, and the children are scored on your own scans. Settings that work together survive together.
1. Scoreon your own scans 2. Selectthe strongest as parents 3. Combineparents into children 4. Varysmall random changes Nextgeneration How a child is made Parent A Parent B Crossover: segments from each parent Child Mutation: one setting changed at random Child Each segment is one setting, such as window or threshold
  1. 01

    Run at the scanner, not only at the hub

    The triage model runs on a small server beside the CT, so the study is scored the moment it is reconstructed, before it is even sent. The flag and the priority travel with the study to the teleradiology hub. If the network link is slow or down, triage still happens locally.

  2. 02

    Baseline by scanner and by hour

    Our head and body CT models are run against the region's own trauma archive, broken down by scanner model and by time of day. The baseline shows where misses happen now: usually the oldest scanner, overnight, on the smallest bleeds.

  3. 03

    Evolve with a recall floor and a time budget

    Candidates are scored on recall for intracranial bleed, cervical spine injury, tension pneumothorax, and solid organ injury, each with its own floor; on false alarm rate per hospital; on worst scanner performance; and on time to score on the local hardware. The trauma and radiology leads choose the operating point from the surviving set.

  4. 04

    Reorder the queue and watch the clock

    Every trauma study enters the worklist with a priority and a time to read target. The most urgent study is at the top, with the finding and its location marked. If a study passes its target unread, the on call radiologist and the trauma team are alerted. The radiologist reads and decides. The model orders the line.

  5. 05

    Learn from every confirmed case

    Operative findings and final reports flow back as ground truth. Each new generation is scored against them, proposed only if it beats the deployed model on every objective, and documented under a predetermined change control plan.

The output is a reordered queue, not a pop up. Left: the worklist as it arrives, in order of time. Right: the same six studies after triage, ranked by the most urgent finding, each with a time to read target.
Before: order of arrival After: order of urgency Routine knee MRI08:12 Chest CT, follow up08:15 Trauma CT, 34 year old08:21 Abdominal ultrasound08:24 Head CT, headache08:30 Trauma CT, 19 year old08:33 Trauma CT, 19 year oldBrain bleednow Trauma CT, 34 year oldOrgan injury10 min Head CT, headacheNo urgent finding1 h Chest CT, follow upNo urgent finding2 h Abdominal ultrasoundNo urgent finding2 h Routine knee MRINo urgent finding4 h
What the evolution changes

How much the model is willing to miss, and how fast it has to be.

These settings pull against each other. More sensitivity means more false alarms unless the preprocessing and the ensemble change with it. We search them as a population.

What we evolveWhy it matters in trauma
Slice handling and reconstruction windowOlder scanners produce thicker slices and more noise. How the model stacks and windows them decides whether a small bleed survives preprocessing.
Per finding sensitivity floorsEach time critical finding gets its own floor. A candidate that trades bleed recall for fracture recall is discarded.
False alarm budget per hospitalA tertiary center with four readers can absorb more flags than a district hospital with one. The threshold is evolved per site.
Study level priority ruleHow the findings in one study combine into a single rank: the worst finding, or a weighted sum. Evolved against the trauma team’s own urgency scale.
Model size and speedScoring must finish on the server beside the scanner within the time it takes to send the study. Speed is an objective, not an afterthought.
Abstention and escalationWhen the model is unsure, the study is marked for a fast human look rather than ranked low.
What you get

The bleed at the top of the list, wherever the crash happened.

Triage at the scanner

Scored the moment the study is reconstructed, with or without a network connection to the hub.

A worklist, not a pile

Studies ranked by urgency with a time to read target, and alerts when the queue falls behind.

Every time critical finding

Head, spine, chest, abdomen, and pelvis read together, ranked by the finding that matters most.

Tuned to your worst scanner

Judged on the oldest machine in the region, so the hospital that needs it most gets the model that works there.

Evidence on every flag

Location, heat map, and confidence, so the radiologist confirms in seconds instead of searching.

Audit trail and change control

Every version logged and reproducible, with a documented path for updates under your device regulation.

Where this fits

Built for the hospital at the end of the road.

If your trauma arrives at hospitals with one CT, one reader, and a long drive to the nearest neurosurgeon, this was built for you.

  • Regional and district hospitals on intercity highways
  • Teleradiology networks covering trauma overnight
  • Health systems with road safety targets they are not on track to meet
  • Trauma systems where transfer decisions depend on a read that is hours away
  • Any setting where the oldest scanner sees the most serious injuries

Trauma triage products were designed in the places with the most radiologists. We design ours for the places with the fewest, measure it on the scanner that performs worst, and set the floor on the finding that kills fastest. It is a harder target. It is also the one that moves the death rate.

Tell us which hospitals your trauma arrives at. We will baseline on your trauma archive by scanner and hour, and show you how many time critical findings move to the top of the queue.
Partner With ALBA