Healthcare

Built for the patients the datasets forgot.

Sickle cell disease and thalassemia cluster in specific populations and regions. The imaging they need is lifelong, subtle, and different from the general case, and the models most hospitals can buy have never seen it. We start from systems that already read the anatomy and evolve them to the cohort.

The problem

A common disease in the wrong place to be rare.

Globally, inherited blood disorders are rare enough that most imaging AI ignores them. In the regions where they concentrate, they are a daily workload.

1 in 6adults carry the sickle cell trait in one high prevalence province, and more than 1 in 100 have the disease
2.6%of newborns in that province are born with sickle cell disease, found by routine screening
73%of deaths from the disease in one hospital series occurred before age 30, led by acute chest syndrome

Sickle cell disease and thalassemia follow geography. They are concentrated in the Middle East, Africa, the Mediterranean, and South Asia, and within those regions in particular provinces and communities where the trait is carried by a large share of the population. A hospital in one of those provinces may see hundreds of children and adults with the disease, each of whom will be imaged dozens of times across a lifetime.

That imaging is specific. A chest X-ray during a pain crisis has to separate acute chest syndrome from pneumonia in the first hours. A child's brain MRI has to find small silent infarcts that predict stroke. A hip MRI has to catch bone death early enough to save the joint. A liver or heart MRI has to measure iron overload in a patient who has had years of transfusions. None of this is what general purpose imaging models were trained to see, and some of it actively confuses them.

Why the standard approach fails

The anatomy is different, and the models were never told.

01

Normal for this patient looks abnormal to the model

Enlarged hearts, thickened skull bones, missing spleens, and bone marrow expansion are routine in this population. A general model flags them as findings and buries the real one in false alarms.

What we do. The model is evolved on confirmed cases from the cohort, so the baseline anatomy of the disease is learned as the baseline, not as pathology.

02

The findings that kill are subtle and fast

Acute chest syndrome begins as a faint new infiltrate on a chest X-ray that already looks busy. The hours between the first film and the diagnosis are where outcomes are decided.

What we do. Recall on early acute chest syndrome is a hard floor in the evolution loop. A candidate that misses it is discarded before anything else is scored.

03

Half the patients are children

Pediatric imaging has its own anatomy, its own protocols, and far less training data. Most imaging AI was validated on adults and says so in the fine print.

What we do. Performance on the pediatric subgroup is a separate objective. The model cannot win on adults by losing on children.

04

No one builds for a market this specific

A product aimed at the global average has no reason to spend a year on one region's disease. So the region's hospitals get the global average.

What we do. We do not build a new product for each cohort. We evolve our existing systems to it, which is a matter of weeks of search, not years of development.

What normal looks like depends on who the model learned from. In this population an enlarged heart, bone changes, and a missing spleen are the baseline. A general model flags all of it. The evolved model flags the one thing that is new.
General model Evolved model Enlarged heart Bone change Absent spleen Marrow change New infiltrate New infiltrate 5 flags, 4 of them false 1 flag, the one that matters
How we solve it

One model family, evolved across a patient’s lifetime.

The same evolutionary loop, applied to each imaging moment the disease creates, from newborn screening to adult complications.

In plain terms

We do not write a new program for this disease. We take models that already know what a chest, a brain, and a hip look like, make dozens of versions of each, and test them on scans from your own patients. The versions that stop raising false alarms on this population and catch the dangerous findings early are kept and combined. A few rounds later, the model treats your patients as the normal case, because for it they are.

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
Newborn and child Brain MRI for silent infarcts EVOLVED READ Small lesions, pediatric floor Pain crisis Chest X-ray within hours EVOLVED READ Chest syndrome vs pneumonia Young adult Hip and shoulder MRI EVOLVED READ Bone death before collapse Transfused patient Liver and heart MRI EVOLVED READ Iron measured, not guessed Every confirmed outcome at any point feeds the next generation for all of them
  1. 01

    Start from a model that already reads the anatomy

    Our imaging systems already read chest X-rays, brain MRI, musculoskeletal MRI, and abdominal studies. We do not start from zero for a cohort. We start from models that know what a lung, a brain, and a femoral head look like, and we evolve what they consider normal and what they consider urgent.

  2. 02

    Baseline on the cohort, by age group

    The models are run against the hospital's own archive of confirmed cases, split by children and adults and by scanner. This is where general models show their false alarm rate on this population. It is usually the first time the hospital has seen that number.

  3. 03

    Evolve with hard floors

    A population of model variants is scored on the objectives the hematology and radiology teams set: recall on early acute chest syndrome, silent infarct detection at small lesion sizes, early bone death staging, iron quantification error, pediatric performance, and false alarm rate on the cohort's baseline anatomy. The floors on the time critical findings are absolute. Everything else is a tradeoff the clinical team chooses from.

  4. 04

    Read inside the clinical pathway

    The evolved models sit inside the hospital's pathways for the disease: the emergency chest X-ray during a crisis, the scheduled brain MRI in childhood, the surveillance MRI for joints and iron. Each flag carries a heat map and a confidence score. The clinician decides.

  5. 05

    Keep evolving with the registry

    Hospitals that treat these patients usually run a registry. Confirmed outcomes from it feed the next generation on an agreed schedule, and every model version is logged and reproducible for the regulator.

Some things are not up for negotiation. For the findings that kill quickly, we set a floor. Any candidate model below it is thrown out before anything else about it is considered, no matter how well it does elsewhere.
Floor: catches 9 in 10 early cases Early cases caught Each dot is one candidate model. Crossed out ones never reach the next generation.
What the evolution changes

What normal means, and how early the model is allowed to commit.

In genetics, when the effect of one gene depends on another, it is called epistasis. The same is true of these settings. We search them together.

In plain terms

The model has to learn two things at once: what is normal for your patients, and how early it is allowed to raise a flag. Those two settings pull against each other. Teach it that an enlarged heart is normal and it may get quieter about everything. Push it to flag early and it may flag the enlarged heart again. We search both at the same time and keep only the combinations that get both right.

What we evolveWhy it matters for this cohort
The reference for normalCardiomegaly, marrow expansion, and an absent spleen are baseline here. The model learns the cohort’s baseline so it stops flagging it.
Weighting of early, time critical positivesEarly acute chest syndrome is faint and fast. The loss weighting is evolved so the model commits on it, under a recall floor it cannot trade away.
Small lesion sensitivitySilent infarcts in children are a few millimeters. Patch size and attention placement decide whether they are seen.
Pediatric and adult branchesWhether to run one model with age conditioning or two specialized members is itself a gene the search decides.
Quantification calibrationIron overload is a number, not a flag. Calibration against the center’s reference measurements is evolved per scanner.
Thresholds per pathwayAn emergency chest film and a scheduled surveillance MRI tolerate different false alarm rates. Each pathway gets its own operating point.
AbstentionWhen the model is unsure, it routes the study to a human without a flag rather than guessing.
What you get

Imaging AI that treats your population as the norm.

Fewer false alarms on your patients

Because the cohort’s anatomy is the baseline, radiologists stop dismissing flags and start trusting them.

Earlier catches on the findings that matter

Acute chest syndrome, silent infarcts, and early bone death flagged at the stage where treatment changes the outcome.

Children covered, not footnoted

Pediatric performance is measured, reported, and protected as its own objective.

Numbers for the transfusion program

Calibrated iron quantification across scanners, so dosing decisions rest on measurement.

In country, in hospital

Deployed on your infrastructure. Patient data does not leave, which satisfies data residency rules by construction.

A documented path through regulation

Every generation logged. A predetermined change control plan so evolution is an approved process, not a resubmission.

Where this fits

The harder the cohort, the better the fit.

This is the version of the problem we built the company for: a real population, a real workload, and no product on the shelf for it.

  • Hospitals and hemoglobinopathy centers in regions with high carrier prevalence
  • Newborn screening programs that now have a cohort to follow for life
  • Transfusion programs that need iron overload monitored across many scanners
  • Pediatric hospitals carrying the stroke prevention workload for the disease
  • Any other inherited or regionally concentrated condition with the same shape: common locally, invisible to global models

Most AI vendors look at a cohort like this and see a small market. We see the clearest possible demonstration of what evolving a model is for. The patients are specific, the findings are specific, the stakes are high, and the standard model is measurably wrong. That is the work we take on first, not last.

Tell us which population your current tools get wrong. We will baseline our models on your cohort, children and adults separately, and show you the false alarm rate before and after evolution.
Partner With ALBA