Home · Research vertical
Medical imaging · live

A consensus engine for chest imaging — built to be right on scans it has never seen.

Fifteen independent, published models read every chest X-ray and vote. We study how to make that ensemble generalise across scanners, sites and populations — through foundation-encoder selection, careful fine-tuning, honest calibration and a discipline of proving (and rejecting) hypotheses. MRI, CT and other X-rays are joining this vertical next.

The reader ensemble

Many validated readers, one calibrated consensus.

Each reader is an independent model with open weights and a measured operating point. They disagree usefully — a finding is surfaced only when readers agree; lone flags are demoted for review, not hidden.

Broad-coverage classifiers Zero-shot open-vocabulary Consolidation specialist Atelectasis specialist Pleural-effusion specialist Pneumothorax (segmentation) Pneumonia (segmentation) Tuberculosis screen Anatomy & cardiothoracic ratio Nodule / mass detector (bounding boxes) Normal / abnormal triage
01

Canonicalize

De-identify & normalize away vendor LUT / windowing shift.

02

Gate

Abstain on non-frontal, wrong-modality or unreadable inputs.

03

Read

Every reader scores independently, with localisation.

04

Consensus

Weighted agreement + per-site calibration + OOD advisory.

05

Review

Calibrated draft for radiologist review & export.

What a read looks like

Every finding, localized and attributed on the image.

The ensemble doesn't just say what — it says where, and who saw it. Attention boxes and trained segmentation masks are drawn back onto the scan; a finding is marked consensus only when readers agree.

TB screen · flagged Nodule · detector 0.71 Consolidation · specialist Pleural effusion · consensus ✓ CTR 0.54
Illustrative composite — a single annotated study rendering the ensemble's localized output. Not a real patient scan.

Reading the annotation

Each box is a claim a reader can defend — and the colour tells you which kind of evidence produced it.

Attention (CAM) localization Trained segmentation mask Consensus — ≥2 readers agree Single reader — flagged for review Quantitative measurement (e.g. CTR)

Localization is normalized back to the original scan and colour-coded by finding category; attention boxes are tightened to the hot core so they mark the finding, not a third of the film.

What we've measured

Fine-tuning & encoder studies, scored the honest way.

Every number below is from our own experiments, scored leave-one-site-out. Specific model and dataset names are withheld; methods are described. Values are research findings, not regulatory claims.

Foundation-encoder bake-off

Cross-cohort generalisation · mean AUC, leave-one-dataset-out (frozen features, identical probe)
Encoder Amedical self-sup · large
0.797
Encoder Bmedical ViT · lightweight
0.756
Encoder Cmedical self-sup · base
0.700
A medically-pretrained, self-supervised encoder is worth several AUC points over a generic one — and choosing the right one is a bigger, cheaper lever than most architecture changes. This picked our deployment encoder.

The model that wins at home can lose away

Supervised-adapted vs self-supervised encoder · AUC
In-domain (home site)
Supervised-adapted
0.818 ▲
Self-supervised
0.785
External (unseen site)
Supervised-adapted
0.815
Self-supervised
0.829 ▲
The supervised encoder led in-domain but lost on a truly external site — its edge was pretraining overlap, not better transfer. Heavier adaptation distorts features to the source; self-supervised features travel. Hypothesis rejected.

Calibration: making confidence mean something

Reliability curve — predicted vs observed probability
predicted probability observed
ideal before (over-confident) after (calibrated)
Per-pathology recalibration cut calibration error ~75× with identical AUC — it fixes the operating point, never the ranking. A stated 80% is now really ~80%.

Diverse normals raise safe rule-out yield

Studies safely auto-cleared at a fixed 98% abnormal-sensitivity
Single-region normals
~24%
+ 3 new geographies
~44%
A small AUC move but a large yield move: adding geographically-diverse confirmed-normal data reshaped the ROC exactly where triage operates, nearly doubling safe auto-clear — advisory only, never autonomous.
The finding-reliability matrix

Which findings are trustworthy — and where they aren't.

The hardest lesson in the whole programme: reliability is finding-specific. Some findings learn cleanly everywhere; some are intrinsically hard; and some collapse only because their report-mined labels are noise, not because the disease is. This matrix is why we build clean-anchored specialists for exactly the red cells.

Radiologist
cohort A
Radiologist
cohort B
Radiologist
cohort C
Report-mined
labels
Pleural effusion
High
High
High
High
Cardiomegaly
High
High
High
Hard
Pneumothorax
High
High
Hard
High
Consolidation
High
Hard
Hard
Noisy
Atelectasis
Hard
Hard
Hard
Noisy
Pneumonia
High
Hard
Noisy
Nodule / mass
Hard
Hard
Hard
Tuberculosis
High
High
Hard
Reliable — learns cleanly from labels here Intrinsically hard / cohort-dependent Label-noise — report-mined labels unreliable not annotated in this cohort

Qualitative summary of our cross-cohort validation; cohorts anonymised. Effusion, cardiomegaly and pneumothorax are genuinely reliable; consolidation and pneumonia are mostly label problems (fixed by adding a clean anchor); atelectasis is near its intrinsic ceiling even with perfect labels.

The hypothesis ledger

What we've proven — and, just as usefully, rejected.

A research group is only as honest as the ideas it's willing to kill. These are hypotheses we tested on held-out data and let the numbers decide.

✓ Proven

The real failure mode is the site, not the model.

In-domain finding-detection AUC ~0.90 collapses to ~0.70 on an unseen scanner — a shift driven by label definitions and acquisition, not the disease.

A classifier can identify the source site from the raw signal with ~98% accuracy; normalization + lung-crop removes that shortcut.
✓ Proven

Add one clean-labelled anchor and noisy sources stop hurting.

Training a reader on report-mined labels alone is the failure; adding a radiologist-labelled anchor to the mix beats noisy-only by up to +0.43 AUC on a clean test.

Deployed as dedicated consolidation, atelectasis and effusion specialists anchored on clean labels.
✓ Proven

Calibration fixes confidence without touching discrimination.

Per-pathology recalibration cut calibration error ~75× while AUC stayed identical — the slope is clamped so recalibration can never re-rank.

Every reader now reports a probability you can act on.
✓ Proven

A self-supervised medical encoder generalises best.

Across three radiologist cohorts, a self-supervised foundation encoder beat a generic one by several AUC points and out-travelled a heavier supervised model externally.

A three-encoder, leave-one-cohort-out bake-off settled the deployment choice.
✕ Rejected

A heavier, supervised-pretrained encoder generalises better.

It won in-domain (+0.033 AUC) but lost on a truly external site (−0.014). The apparent edge was pretraining overlap, not transfer.

Kept on the shelf, unused. Self-supervised features travel further.
✕ Rejected

Single-site full fine-tuning is the way to adapt.

Full fine-tuning gains a little in-domain but loses out-of-distribution — it distorts pretrained features to the source site.

We adapt with frozen features, light probes or LoRA, and score leave-one-site-out.
✕ Rejected

More models in the ensemble always help; so does flip-TTA.

A larger cross-domain ensemble lowered in-domain AUC, and horizontal-flip test-time augmentation regressed laterality-dependent findings.

Both off by default; calibration and clean labels mattered far more than brute ensembling.
✕ Rejected

Conditioning cardiomegaly on view (AP) fixes false positives.

The assumed view-conditioning fix didn't hold on our data; the real driver was a flat cardiothoracic-ratio threshold over-calling ~44% of screening films.

Fixed with a tuned threshold + lateral suppression instead — measure before you rebuild.
Coming to this vertical

MRI, CT and other X-rays — same method, new modalities.

The generalization gap isn't unique to chest radiographs. We're extending the ensemble to published, open-weight models for other modalities, carrying the same discipline: adapt lightly, validate leave-one-site-out, calibrate per site, keep only what survives.

CT

Volumetric chest CT — nodule characterisation and risk models, evaluated for cross-scanner robustness before anything is adopted.

MRI

Sequence-aware models where acquisition shift is even larger than in X-ray — a demanding test of our normalization-first approach.

Other radiographs

Musculoskeletal and abdominal projections, reusing the consensus, calibration and abstention machinery already validated on chest.

Bring a modality or a cohort →

Have imaging data from a scanner we've never seen?

That's exactly the test our method is built for — and the best kind of collaboration.