DualRep: Barrett's neoplasia detection in a low-prevalence setting

Submission to the RARE26 challenge (EndoVis, MICCAI 2026). The task is binary classification of endoscopy frames as Barrett's neoplasia versus non-dysplastic Barrett's oesophagus (NDBE).

DualRep is a two-arm ensemble. What makes it unusual is what the arms differ in: pretraining corpus and architecture, rather than training recipe. That choice came out of a failure. Our previous submission ensembled fifteen checkpoints over one backbone with three different recipes, and it barely moved. Members that share a visual representation fail on the same images, and on this metric only the failures matter.

---

The metric decides the design

Submissions are scored by positive predictive value at 90% recall, under a simulated 100:1 negative-to-positive prior:

PPV@90R = 0.90 / (0.90 + 100 × FPR@90R)

It is a strictly monotone function of a single number: the false-positive rate at the threshold that achieves 90% recall. Chance sits at 0.0099.

Two consequences shaped everything that follows.

Only ordering matters. Any monotone rescaling of scores leaves the metric untouched. So calibration and pooling can help only by changing which images outrank which.

The threshold is set by one image. At 90% recall on roughly 103 development positives, the operating point is the 10th-percentile positive. We once measured two poolings that correlated at Spearman 0.99997 and still differed 2.5× in FPR@90R. Treat any single FPR@90R delta as noise until it survives a group-clustered bootstrap.

---

The system

flowchart LR
    IMG["Endoscopy frame"] --> P0["P0 preprocessing<br/>resize + ImageNet normalise"]

    P0 --> A["Arm A · DINOv2 ViT-B/14 reg4<br/>GastroNet-5M self-supervised<br/>LoRA r16 a32, last 6 blocks<br/>336 px"]
    P0 --> B["Arm B · ResNet50<br/>Billion-Scale-SWSL + GastroNet-5M<br/>layer4 + fc trainable<br/>384 px"]

    A --> CA["Platt a,b per checkpoint<br/>fitted on its own out-of-fold data"]
    B --> CB["Platt a,b per checkpoint<br/>fitted on its own out-of-fold data"]

    CA --> MA["mean of 5 folds"]
    CB --> MB["mean of 5 folds"]

    MA --> F["equal-weight mean<br/>across arms"]
    MB --> F
    F --> OUT["neoplasia probability"]

Loading

	Arm A	Arm B
backbone	DINOv2 ViT-B/14 (reg4)	ResNet50
initialisation	GastroNet-5M, self-supervised	Billion-Scale-SWSL + GastroNet-5M
adaptation	LoRA r16/α32 on attn.qkv + attn.proj, last 6 of 12 blocks	layer4 + fc; trunk frozen in eval()
resolution	336	384
trainable	0.443 M of 86.42 M (0.51%)	14.97 M of 23.51 M

Why it is built this way

Two arms, not one backbone in several recipes. The fifteen-checkpoint version cut in-distribution FPR@90R by 39% and moved the leaderboard by 1.2 points. Recipe diversity does not decorrelate errors. These two arms correlate at Spearman 0.48 to 0.59 and share only about half their bottom-decile positives, which are exactly the images that set the threshold.

Probabilities, calibrated before averaging. The arms are different architectures trained on different corpora, so their logit scales are not comparable. Fitted Platt slopes span 0.46 to 1.19 and intercepts run from −3.25 to +1.98. Averaging raw scores actually measured negative on both held-out centres (center_2 AUPRC −0.0187). Each checkpoint therefore carries an affine calibration fitted on its own out-of-fold predictions, data that checkpoint never trained on, frozen at build time.

No rank fusion. Percentile ranks would be computed within whatever stack the container is handed, while the metric is pooled over the whole test set. That would make every prediction depend on its batch's composition. Frozen offline calibration does not.

Equal weights, by predeclaration rather than search. With only two usable centres, any fitted fusion weight gets selected on the very data that would validate it. That is precisely what put +0.12 of pure optimism into an earlier fusion.

What was submitted

image	sha256:6571524e410b3e59a0e7af1b66b433495c2f7714c71439def0f74a8af88d376c
tag	rare26-twoarm-probavg
archive	twoarm_10model_probavg.tar.gz, 6,357,006,153 bytes
runtime	6.1 ms per frame per model, plus about 25 s fixed start-up (RTX 5060 Laptop)

Verified under --network none --gpus all: both arms load their five checkpoints, calibration executes, and fixture scores carry no ties and no saturation. Every model is built with pretrained=False / weights=None and loaded from checkpoints baked into the image, so nothing touches the network at inference. All state dicts load with strict=True.

---

Results

Leave-one-centre-out: train on the remaining centres, then evaluate on a centre no model has seen. Paired bootstrap clustered on near-duplicate groups.

held-out centre	AUROC	AUPRC	FPR@90R
center_1 · 2,279 images, 61 positives (2.68%)	0.9870	0.9012	0.0140
center_2 · 816 images, 97 positives (11.9%)	0.9930	0.9754	0.0014

Read these as a different regime from the leaderboard, not merely as optimistic numbers. Earlier configurations from this project measured FPR@90R near 60% on the development leaderboard, against low single-digit percentages on a held-out centre here. That gap spans one to two orders of magnitude, and it is a property of the evaluation rather than of any particular model: one held-out centre is a far easier problem than twelve heterogeneous ones.

Two caveats matter for reading the table. The 60% figure comes from those earlier, externally evaluated configurations, and does not correspond to the rows above. And the exact configuration submitted here, plain DualRep with no prior shift and no noisy-OR, was never scored on the development leaderboard at all. So treat this table as valid for ranking two candidates against each other, and invalid as an absolute expectation. EXPERIMENT_LOG.md §1 documents the calibration in full.

---

What actually worked

EXPERIMENT_LOG.md is the honest account: twenty-seven levers tested, three ever worked. All three changed what the model represents. Nothing that changed only how it trains has survived cross-centre transfer here.

1. The endoscopy-pretrained backbone. This is the single largest effect measured in the project. Swap stock DINOv2 (LVD-142M) for the same architecture self-supervised on GastroNet-5M:

leave-one-centre-out AUPRC	stock	domain
center_1	0.634	0.873
center_2	0.863	0.975

Confidence intervals do not overlap on either centre. A linear probe on frozen features, with no LoRA at all, confirms it independently: 0.521 → 0.695 and 0.763 → 0.938.

2. Two-arm representation-diverse fusion. The system above. All six deltas across both centres and all three metrics came out positive, which no single-backbone ensemble achieved.

3. Jigsaw as a third arm. Useful for error diversity, not as a standalone recipe.

What did not work, each of these measured rather than assumed: test-time augmentation, seed ensembling, threshold calibration, deeper LoRA, resolution in either direction, a fine-tuned ResNet arm, centre-balanced loss, positive-enriched batching, a capped tail loss, metric-aligned checkpoint selection, centre-invariant scores, patch-token MIL, a PPV@90R surrogate loss, acquisition-conditional tail normalisation, MaxViT at every point on its capacity curve, stock DINOv3 in every adapter configuration, and a full segmentation and localisation route built out to a working decoder (val Dice 0.667). Three ingredients transplanted directly from the RARE25 winner's published code all lost.

The most expensive lesson was methodological. Prior shift plus noisy-OR looked decisive locally, a 22.6% cut in FPR@90R on leave-one-centre-out. We shipped it. It failed externally, costing 0.024 AUPRC and delivering none of the tail gain it was supposed to buy, so it has been removed. The submitted container is the plain configuration that passed the predeclared gate before any tail machinery was layered on top.

The deeper problem is that the local validator cannot see the failure mode at all. A 36-condition acquisition-shift stress sweep bottoms out at AUROC 0.9669 while the leaderboard runs at 0.82. No perturbation available to us gets within 0.15 AUROC of the scored regime.

---

Reproduction

You need Python 3.12, a CUDA GPU, and the challenge data.

Install in two steps. PyTorch must come from the CUDA 12.8 index, because the shipped checkpoints were written by torch 2.11.0+cu128 and that build is the one verified to carry sm_120 kernels, which Blackwell cards require. A default pip install torch will give you a CPU or non-cu128 wheel and fail at load time.

python -m venv .venv
.venv/Scripts/python -m pip install torch==2.11.0 torchvision==0.26.0 \
    --index-url https://download.pytorch.org/whl/cu128
.venv/Scripts/python -m pip install -r requirements.txt

On Linux or macOS the interpreter path is .venv/bin/python instead.

flowchart TB
    D1["RARE26 training release<br/>3,095 images · 158 neoplasia"]
    D2["EndoVis 2015 Barrett's set<br/>100 images · 50 neoplasia"]

    D2 --> E["ingest_evc.py"]
    D1 --> FLD
    E --> FLD["make_folds.py + build_combined.py<br/>folds_v2.csv · grouped 5-fold"]

    FLD --> T1["train_dinov2.py<br/>x 5 folds"]
    FLD --> T2["train_resnet.py<br/>x 5 folds"]

    T1 --> O1["best.pt<br/>oof_fold_k.csv"]
    T2 --> O2["best.pt<br/>oof_fold_k.csv"]

    O1 --> PL["fit_platt.py<br/>platt.json"]
    O2 --> PL
    O1 --> RES["stage into resources/"]
    O2 --> RES

    PL --> BLD["docker build"]
    RES --> BLD
    BLD --> TAR["docker save + gzip<br/>twoarm_10model_probavg.tar.gz"]

    FLD --> LOCO["train_cross_centre.py<br/>compare_loco.py"]
    LOCO -.-> T1

Loading

1. Data

Place the challenge training release at RARE25-train-data/<centre>/<ndbe|neo>/<image_id>.png, with its inventory at data_inventory_rare25.csv. That file ships with the challenge data and is not generated here. Fold grouping reads its duplicate_group column.

The third centre (evc, the EndoVis 2015 Barrett's set) is optional. It goes into the training pool and is excluded from every reported metric, because both stock and domain models score about 1.0 AUROC on it without training on it. It cannot discriminate between candidates.

python data_split_scripts/ingest_evc.py     # -> evc_inventory.csv (optional third domain)

Neither the data nor the pretrained weights are redistributed here.

2. Pretrained backbones

Download from Theta Vision Cortex into the project root, and verify checksums against docs/MODEL_PROVENANCE.md before use. The loaders assert matched key counts, because strict=False will silently hand you a random backbone.

• dinov2.pth for GastroNet-5M DINOv2 ViT-B/14 reg4
• RN50_Billion-Scale-SWSL%2BGastroNet-5M_DINOv1.pth

3. Folds

python data_split_scripts/make_folds.py --out folds_v1.csv
python data_split_scripts/build_combined.py --folds folds_v1.csv --out folds_v2.csv

folds_v2.csv is the split every shipped model uses: 3,195 images and 208 neoplasia across five folds of 638 to 640 images, each holding 41 or 42 positives, grouped so no group_id spans a fold boundary. Challenge images are grouped by near-duplicate clusters and EVC images by patient ID. make_folds.py asserts that invariant.

The distinction matters when reading the numbers. The challenge release ships no patient, exam or video identifiers, so grouping for center_1 and center_2 is by perceptual-hash near-duplicate cluster, which is a proxy for patient identity and not a substitute for it. Only EVC carries true patient IDs. Every bootstrap in this repository is therefore clustered by group, not by patient.

4. Train the ten shipped checkpoints

for k in 0 1 2 3 4; do
  python training/train_dinov2.py --folds folds_v2.csv --fold $k \
      --init dinov2.pth --out runs/V2_dinogn_fold$k
  python training/train_resnet.py --folds folds_v2.csv --fold $k --arch resnet \
      --init 'RN50_Billion-Scale-SWSL%2BGastroNet-5M_DINOv1.pth' \
      --freeze-until layer3 --out runs/V7_swsl_fold$k
done

Each run writes best.pt and oof_fold<k>.csv. Everything else is defaults, recorded in each checkpoint's args: batch 32, 30 epochs, lr 5e-4 (DINOv2) or 1e-4 (ResNet), weight decay 1e-4, AdamW, cosine schedule, BCEWithLogitsLoss with a global pos_weight, P0 preprocessing, no augmentation, seed 42, checkpoint selected on inner-split AUPRC with patience 7.

Check GPU memory before launching on new hardware. An oversized batch does not OOM on Windows. It spills to system RAM over PCIe and runs up to 32× slower, in complete silence. See EXPERIMENT_LOG.md §5.

5. Evaluate

# leave-one-centre-out, the protocol every decision in this project was made on
python training/train_cross_centre.py --folds folds_v2.csv --arch dinov2 \
    --holdout center_1 --init dinov2.pth --tag mytag

# paired A/B of two LOCO runs on the same held-out centre
python evaluation/compare_loco.py \
    --baseline runs/LOCO_dinov2_rest_to_center_1_domain \
    --candidate runs/LOCO_dinov2_rest_to_center_1_mytag --folds folds_v2.csv

Do not select on pooled or per-fold cross-validation. It produced four documented wrong conclusions on this project, and EXPERIMENT_LOG.md §4 lists every one of them.

6. Build the container

# stage the ten checkpoints under the names inference.py globs for
mkdir -p submission_template/resources
for k in 0 1 2 3 4; do
  cp runs/V2_dinogn_fold$k/best.pt submission_template/resources/dinov2_fold$k.pt
  cp runs/V7_swsl_fold$k/best.pt   submission_template/resources/swsl_fold$k.pt
done

# refit the per-checkpoint calibration from the out-of-fold predictions of step 4
python evaluation/fit_platt.py \
    --arm dinov2=runs/V2_dinogn_fold{k} \
    --arm swsl=runs/V7_swsl_fold{k} \
    --out submission_template/platt.json

docker build -t rare26-twoarm-probavg submission_template
docker save rare26-twoarm-probavg | gzip -c > twoarm_10model_probavg.tar.gz

platt.json holds [a, b, prior_fit] per checkpoint. The copy committed here is the one the submitted container shipped, and fit_platt.py regenerates it exactly. inference.py raises if any checkpoint is missing its entry, so a forgotten copy fails loudly instead of quietly shipping uncalibrated scores.

do_build.sh and do_save.sh are the organizers' template scripts, kept unmodified. Prefer the two commands above. do_save.sh rebuilds before saving, can hit a Docker context lock, and tags the image example-algorithm-closed-testing-phase.

Verify offline before uploading:

docker run --rm --network none --gpus all \
    -v /path/to/input:/input -v /path/to/output:/output rare26-twoarm-probavg

---

Repository layout

path	contents
data_split_scripts/	fold construction and third-domain ingestion
preprocessing/	preprocessing variants (P0 to P4) and augmentation pipelines
training/	fold trainers, the leave-one-centre-out trainer, the segmentation decoder
evaluation/	the metric, paired A/B testing, frozen probes, feature banks, duplicate audit, shift stress, calibration fitting
submission_template/	the container: Dockerfile, inference.py, model/, platt.json
submission_probe/	single-model probe container
docs/MODEL_PROVENANCE.md	every weight file's source, access date, checksum and licence status
THIRD_PARTY_NOTICES.md	third-party code, weights and datasets, and their terms
EXPERIMENT_LOG.md	the complete experimental record, every lever tested, with numbers

Checkpoints and challenge data are excluded. Stage them as described above.

---

Licence and third-party assets

This repository's original code is released under the MIT License (see LICENSE). The submission_template/ directory is derived from the organizers' RARE submission template and retains its original license in submission_template/LICENSE. Third-party datasets and pretrained weights are not redistributed and remain subject to their respective licenses.

THIRD_PARTY_NOTICES.md lists each third-party item and its terms. docs/MODEL_PROVENANCE.md records every weight file's source, access date, checksum and licensing status.

No private or non-public external data was used at any point.