Healthcare

Every scan in the network, read twice.

A hospital network has one shared problem: the scans that most need an expert are taken where the experts are not. We put a second read on every CT, MRI, and ultrasound in the network, evolved to the hospitals at the edge, so a cancer found in a regional hospital is found as early as one in the capital.

The problem

Imaging volume grows faster than the people who read it.

Fast growing health systems add scanners in months and specialists in decades. The gap lands on the regions.

173%rise in breast cancer incidence over fifteen years in one national registry, with half of all cases diagnosed before age 50
1 in 2breast cancers in that registry found at a regional or distant stage, with the worst stage mix in peripheral regions
2 of 3radiologists in the same system concentrated in three cities, and most of them trained abroad

Teleradiology networks were built to close this gap. A regional hospital sends its scans to a central hub, where a radiologist reads them remotely. It works, and it scales the people. It does not scale their attention. Every scan is still one pair of eyes under time pressure, and the scans from the periphery arrive with the hardest cases: later stage disease, more variety in equipment and protocol, and patients who look nothing like the textbook.

Off the shelf imaging AI does not close the gap either. Most products read one finding on one modality, usually chest X-ray, and were validated on populations that are older, less dense, and more uniform than the ones a growing network serves. A product that is excellent on average can be quietly wrong on the exact hospitals the network bought it for.

In plain terms

A teleradiology network works like a call center for scans. A small hospital takes the image and sends it to a central office where a specialist reads it. That solves the problem of having no specialist in town. It does not solve the problem of one specialist reading two hundred scans in a shift, or of the hardest scans arriving from the places with the least support.

The difference between early and late is a few millimeters. A model trained on an archive full of large, obvious tumors learns to wait for them. We evolve ours on confirmed early cases so it commits while the finding is still small and the treatment still simple.
Time Later Evolved model flags here 6 mm Standard model flags here 20 mm
Why the standard approach fails

The hard cases are the point, and they are where trained models stop.

01

Younger patients, denser tissue

When half the breast cancers are in women under 50, the images are denser and early findings are subtler. Models trained on screening populations in their sixties have not seen enough of this.

What we do. The model is evolved against confirmed early cases from the network itself, with a hard floor on recall for dense tissue that the search cannot trade away.

02

Twenty scanners, twenty looks

A network mixes vendors, protocols, and generations of equipment. A model fit to the hub's own scanner degrades on the regional one that matters most.

What we do. Every candidate model is scored per site and per scanner, and judged by its worst site, not its average.

03

Training data skewed to late disease

Where cancer presents late, the archive is full of large, obvious findings. A model trained on it learns to wait for them.

What we do. Early cases are weighted and augmented in the evolution loop so the population of models learns to commit before the finding is large.

04

One finding per product

Buying a separate tool for breast, lung, liver, and colon means four vendors, four integrations, and still no coverage on CT, MRI, or ultrasound.

What we do. One platform reads the three modalities for the cancers that matter in your population, with one integration and one audit trail.

Judged by the weakest hospital, not the average. A model can score well on average and still fail the one site that needed it most. We score every candidate per site and rank it by its worst one.
92 90 Capital hospital 90 89 City hospital 88 88 Regional hospital 61 86 Regional hospital 70 87 District clinic Standard modelEvolved modelLevel you need

Illustrative. Scores are accuracy on each hospital’s own scans.

How we solve it

A second read inside the hub, evolved to the edge.

Deployed on the network's own infrastructure. Scans never leave the country, and the model learns from the hospitals it serves.

In plain terms

Think of it like breeding for a trait. We make dozens of versions of the model, test every one on scans from your own hospitals, keep the versions that catch the most early cancers without crying wolf, combine them, and repeat. After a few rounds, the model that comes out was shaped by your hospitals, not by a dataset from somewhere else.

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
Regional hospital Regional hospital District clinic Tertiary center Mobile unit Teleradiology hub ALBA second read CT · MRI · Ultrasound Evolved per site Prioritized worklistUrgent findings read first Flag with evidenceHeat map, region, confidence Discordance reviewWhen reader and model disagree Confirmed outcomes feed the next generation
  1. 01

    Deploy inside the network

    The system runs on the hub's own servers or an in country cloud. It connects to the PACS and the teleradiology queue. No image and no report leaves the network, which satisfies data residency rules by construction.

  2. 02

    Baseline by site and scanner

    Our trained imaging models are run against a held out slice of the network's own archive, broken down by hospital and by scanner. Both sides see exactly where the starting point is strong and where it is not.

  3. 03

    Evolve to the edge

    We keep a population of model variants alive and score each one on the objectives the network sets: recall for early breast, colorectal, lung, and liver findings at a fixed false positive budget, performance on the single worst site, the gap between patient groups, calibration, and speed. The strongest survive, recombine, and improve. The result is a set of options with different tradeoffs, and radiology leadership picks the operating point.

  4. 04

    Second read on every study

    Every CT, MRI, and ultrasound gets read by the model as it lands. Urgent findings move to the top of the worklist. Each flag carries a heat map, the region, and a confidence score. Where the radiologist and the model disagree, the case is routed to a second radiologist. The reader decides; the model never signs a report.

  5. 05

    Keep evolving after launch

    Pathology confirmed outcomes flow back into the evaluation set on an agreed schedule. When a new generation beats the deployed model on every objective, it is proposed, reviewed, and documented under a predetermined change control plan. Every version is reproducible and every change is logged.

One scan, two reads, and a rule for disagreement. The radiologist reads as usual. The model reads the same scan. When they agree, the report goes out. When they disagree, a second radiologist looks before anything is signed. The model never signs a report.
The scan Radiologist reads Model reads Agree? Reportsigned by the radiologist Second radiologistlooks before signing Yes No
What the evolution changes

The settings that decide whether an early cancer is caught.

These interact. The right threshold depends on the loss weighting, which depends on the preprocessing. That is why we search them as a population, not one at a time.

In plain terms

A model has hundreds of dials. Turning one changes what the others should be set to, the way adjusting the focus on a camera changes the right exposure. Setting them one at a time gets you stuck. Testing many combinations at once, and keeping the ones that work together, is how you find the setting that catches a six millimeter nodule instead of a twenty millimeter one.

What we evolveWhy it matters in a network
Intensity windows and preprocessingDifferent scanners and protocols produce different images of the same tissue. The right normalization is per site, not global.
Patch size and attention placementEarly findings are small. Where the model looks, and at what scale, decides whether it sees a 6 mm nodule or a 20 mm one.
Weighting of early and rare positivesCounteracts an archive skewed toward late disease, so the model learns to commit on subtle findings.
Augmentation policyStretches scarce early cases without inventing findings. The strength is evolved, not fixed, because it depends on how rare the finding is.
Decision thresholds per finding and per siteA false positive budget that fits a tertiary center overwhelms a district clinic. Thresholds are tuned to the reading capacity at each site.
Ensemble membership and votingWhich model variants vote, and how their votes are weighted, can be conditioned on the scanner that produced the study.
Abstention rulesLow confidence cases route to a human without a flag, so the system knows what it does not know.
What you get

One platform, three modalities, every hospital in the network.

Coverage, not a single product. Most imaging AI a network can buy reads one finding on one modality. One platform covers the cancers that matter in your population across the scans you already take.
Typical imaging AI Breast Lung Liver Colon Chest X-ray CT MRI Ultrasound ALBA Breast Lung Liver Colon Chest X-ray CT MRI Ultrasound Illustrative coverage. The exact set is chosen with your clinical team.
In plain terms

You are not buying a program that has to be rewritten for your hospitals. You are buying a program that already works and a process that reshapes it to your hospitals, measured on your own scans, with your own radiologists choosing how cautious it should be. The program keeps improving as your results come in, and every version of it is on record.

Coverage where there was none

Second reads on CT, MRI, and ultrasound for the cancers that matter in your population. Not only chest X-ray, and not one finding per vendor.

Equal reads at the edge

The regional hospital gets the same quality of second read as the capital, because the model was judged on the regional hospital.

Evidence on every flag

A heat map, the region, and a calibrated confidence score. Radiologists see why, not just what. Discordant cases get a second human read.

Regulation built in

In country deployment, a full audit trail of every model version, and a predetermined change control plan so that evolution is a documented process rather than a series of resubmissions.

Dashboards per site

Recall on confirmed cases, turnaround time, flag rate, and discordance rate for every hospital and scanner, updated as outcomes come in.

A system that keeps improving

Each confirmed outcome makes the next generation better. The network's own data becomes its advantage, and it stays inside the network.

Where this fits

Built for the hard version of the problem.

If your imaging network is uneven, your population is young, or your cancers present late, this was built for you.

  • National or regional teleradiology networks with hundreds of connected sites
  • Hospital groups where specialist staffing varies widely between locations
  • Health systems whose cancers present at a later stage in outlying regions
  • Systems standing up or scaling screening programs for breast and colorectal cancer
  • Populations where disease strikes younger than the datasets most models were trained on

Most imaging AI was built to be good on the common case and sold the same way everywhere. We built ours to be evolved to the uncommon case, and we measure it on the hospital where it is hardest. If a model is going to be wrong somewhere in your network, we would rather find that site in the baseline than have you find it in a patient.

Bring us the hospitals where your reads are hardest. We will baseline our models on your archive, site by site, and show you what the evolved version finds that the standard one misses.
Partner With ALBA