Layer 8 Labs research · CP-HRIC
Reading leukaemia
from a photograph
of a cell.
CP-HRIC (Coordinate-Preserving High-Resolution Image Cytometry) is our research into AI that looks at ordinary stained blood cells and aims to recover much of what today takes an expensive flow cytometer or a genetic test.
every cell keeps
its address
The idea
Put the cells on a slide.
Let software do the rest.
A flow cytometer pushes cells past lasers one at a time and measures them once. It is powerful, and it costs anywhere from tens to hundreds of thousands of dollars before service, reagents and trained staff. We think a camera and a model can recover much of that information.
Commodity hardware
Cameras, optics, motors and GPUs are made at enormous scale. The intelligence lives in software, and a trained model costs almost nothing to copy.
Interpretation anywhere
Once a slide is an image it is just data. The analysis can run on a local machine, at a regional hospital or in a data centre.
Every cell stays addressable
A photographed cell is still on the slide, at a known position. The cells the model flags can be looked at again, re-imaged or sent for molecular confirmation.
How the model works
Two stages, one answer per patient.
Cell encoder
A neural network turns each single-cell image into 12,800 measurements. We train it ourselves on labelled cells from public datasets, so it learns what matters in a blood cell.
220,960 labelled training cells
Patient classifier
A reader model looks at every cell from one patient and gives a patient-level answer, pointing to the cells that drove it. The models we publish are multi-task models: each answers several questions with one model.
65 cell types learned, from 5 public datasets
it shows us which
cells mattered
What it does today
Genetic subtype, from an ordinary blood smear.
Give the model one patient’s blood-smear cells and it identifies which of these it is looking at. The first four are genetic subtypes of acute myeloid leukaemia (AML) that are normally identified by molecular testing.
- PML::RARAAcute promyelocytic leukaemia (APL)
- NPM1NPM1-mutated AML
- CBFB::MYH11Core-binding-factor AML
- RUNX1::RUNX1T1Core-binding-factor AML
- ControlHealthy donor
These five are Experiment 001. Experiment 002 asks the APL question again at a second hospital, and our multi-task models answer both questions from one model, scored on each experiment’s own test (see Experiments below). The cell encoder was trained entirely on public data. The published study it set out to replicate (Hehr et al. 2023, PLOS Digital Health) relied on a private archive of more than 300,000 cells.
Against the papers
Every study we set out to match, scored its own way.
One row per published study, measured exactly the way that study measured its headline claim, on its own patients. The green line is our CP-HRIC-MIL multi-task model and the orange line under it is our CP-HRIC-S1 multi-task model: each is the one model that answers every question, so both are scored on every study. The dashed line is the paper’s number.
Bar = mean across seeds; whisker = ± the spread across seeds; ◆ = those seeds’ models combined into one. Beats = our mean is above the paper’s number by more than run-to-run noise (twice that model’s measured run-to-run spread on that test, or a conservative default where it has not been measured yet, divided by √seeds). Matched within margin of error = above the paper, but by less than run-to-run noise. Matched = our range overlaps it. Approaching = within 0.05. In progress = further. Preliminary = fewer than 10 seeds. The axis runs from 0 to 1.
- +0.275macro-F1 from our first build (0.552) to v3e (0.827), 1 run each
- 189patients in the public benchmark
- 100%public training data
Macro-F1 averages how well each of the five answers is recognised (1.0 is perfect; random guessing scores 0.20). Scored on the published study’s own patient-level folds, so every patient is tested exactly once.
Goal
Two models, one goal: beyond the state of the art.
CP-HRIC-MIL and CP-HRIC-S1 take different approaches to the same questions. Both are measured against every paper's own test, and both are being developed to go past the current models. Datasets are added one at a time, and every earlier answer is re-scored after each addition.
One row per published study: what it claims, the data it used, and the number it reported, measured its own way. Beside it, what our multi-task models score on that same test. Each is ONE model that answers every question it has learned, so the same models appear on every row. DinoBloom, a published foundation model for blood-cell images, is shown last, in grey, as a reference.
| Paper | Question and Data | Published Results | CP-HRIC-MILmulti-task model | CP-HRIC-S1multi-task model | DinoBloompublished foundation model · reference |
|---|---|---|---|---|---|
| SCEMILAHehr 2023 | An AI reading the white cells on a blood smear can name the genetic subtype of AML (five subtypes).Munich (TCIA), 189 patients | macro-F1 over the five subtypes, the paper’s five patient-level folds0.81 | 0.841 ± 0.01910 seeds+0.031 vs 0.81Beats | 0.809 ± 0.01610 seeds−0.001 vs 0.81Matched | 0.834 ± 0.013110 seedsreference, no deltaDinoBloom-B + an attention reader |
| MILLIEManescu 2023 | An AI can tell APL, the emergency form of AML, from other AML using blood-smear cells alone.Johns Hopkins, 106 patients | AUROC 3-fold cross-validation0.96 ± 0.02 | 0.926 ± 0.03110 seeds−0.034 vs 0.96Matched | 0.885 ± 0.02510 seeds−0.075 vs 0.96In progress | 0.967 ± 0.014210 seeds+0.007 vs 0.96DinoBloom-G + MILLIE-style head |
| DeepAPLSidhom 2021 | Deep learning on blood-smear cells picks out APL in patients it never saw in training.Johns Hopkins, 82 train / 24 held out | AUROC on 24 held-out patients0.86score matched; model not re-run | 0.869 ± 0.03410 seeds+0.009 vs 0.86Matched within margin of error | 0.864 ± 0.03810 seeds+0.004 vs 0.86Matched within margin of error | 0.889 ± 0.009210 seeds+0.029 vs 0.86DinoBloom-G + MILLIE-style head |
| APL in marrowEckardt 2022 Planned | APL from bone-marrow smearsDresden | AUROC0.86 | not yet run | not yet run | not yet run |
| LeukemiaAttriITU Planned | Per-cell morphology attributesITU | No comparable headline: a dataset paper whose baselines score cell detection (mAP), not cell-type classification. | not yet run | not yet run | not yet run |
| Marrow atlasBandyopadhyay 2024 Planned | Marrow niches, multiplex imagingPenn / CHOP, CODEX imaging | No comparable headline: a descriptive atlas with no held-out prediction to match. | not yet run | not yet run | not yet run |
- DinoBloom’s pretraining included the TCIA (Munich) images these SCEMILA patients come from (confirmed), so its SCEMILA score is a reference with no delta.
- Johns Hopkins images were possibly included in DinoBloom’s pretraining, per DinoBloom’s documentation: its paper’s list of pretraining datasets omits them, but its GitHub README lists them.
Beats = our mean is above the paper’s number by more than run-to-run noise (twice that model’s measured run-to-run spread on that test, divided by √seeds; judged for the two joint models only) · Matched within margin of error = above the paper, but by less than run-to-run noise · Matched = our mean ± the spread across training runs overlaps the paper’s reported range (or reaches its number when it gives no range) · Approaching = within 0.05 · In progress = further away.
Each result cell: the mean across independently trained seeds ± the spread across seeds, the number of seeds, the difference from the paper’s number, and the result chip. Fewer than 10 seeds = preliminary. The chip takes its column’s colour (green = cp-hric-mil, orange = cp-hric-s1); the word, and how solid the chip is, give the result. DeepAPL’s number is its published score; its model was not re-run.
Every experiment, one page each →What each model is, and how System One works →
Where it’s going
From five answers to one model that reads the whole slide.
- Next
Set the benchmark
Move past the leading public foundation model for blood-cell images on this benchmark, with our own encoders and with that model fine-tuned on our data.
- Next
Travel well
Perform just as strongly on patients from other hospitals, scanners and stains. We already score our models on a second hospital’s patients, and one of our multi-task models already reaches DeepAPL’s published score on its 24 held-out Johns Hopkins patients (Experiment 002).
- Started
One model, many answers
Our multi-task models already answer two questions with one model. Next, from one unlabelled photo: find every cell, name its type and morphology, flag unusual or malignant cells, and answer patient-level questions across AML subtypes, APL, childhood leukaemias and MDS.
The long view
Imaging does the broad search.
Biology goes where it matters.
A scanner photographs hundreds of thousands of cells, the model reads every one, and the few that need a closer look are found again on the slide for targeted imaging or molecular confirmation. Along the way: more leukaemia subtypes, including KMT2A-rearranged disease, a broader haematology foundation model, and hardware that can return to any flagged cell.