Mammo-FM risk checkpoints: BU S1 KD + S2
This release contains five BU fold-specific final Stage 2 risk models
initialized through Stage 1 MIRAI knowledge distillation. It is the
historical all_years/aug_n/from_stage1 model family, not Stage 1 alone and
not the S2-only comparison variant. These checkpoints replace the previously
released Mayo-labelled Stage 1/Stage 2 pair, which was not the intended BU
benchmark release.
Download and provenance
All five files contain 184,831,762 bytes. Stored epoch values are reproduced without changing their indexing convention. The release manifest contains the same checkpoint identities and hashes.
| BU fold | Final S1 KD + S2 checkpoint | Stored epoch | SHA-256 |
|---|---|---|---|
| 1 | Download fold 1 | 2 | f57730d674716c5870e5f0678c82c2a21cdff37eea518059f09809032830653e |
| 2 | Download fold 2 | 4 | cac0884effa796de7fc32ac98b630301755d1d72a12a52addab24853a2c618a9 |
| 3 | Download fold 3 | 6 | f6b1cf5845c708a530ef8f5f98ecf2db87b54217f74b430e9b068ec99a40d1d6 |
| 4 | Download fold 4 | 6 | b1bf8769f3d6ee86253e36a473047396fb2e9d34470c440b8c5c4421dcbb9742 |
| 5 | Download fold 5 | 5 | 0df00131c5a91ccd2e9d464bdbfb561f9604d589e38e61fa2903de5c50b8490f |
hf download batmanLab/Mammo-FM \
--include 'risk_checkpoints/BU_S1_KD_S2/*' \
--local-dir /path/to/checkpoints
The weights are released byte-for-byte from the historical best-validation
checkpoints. Their payloads contain model, optimizer, scheduler, and
epoch; fold and training-stage lineage come from the corresponding original
experiment artifacts. No images, reports, patient-level tables, predictions,
feature caches, or API credentials are included. Store downloaded weights
and all experiment outputs outside the source repository.
BU held-out evaluation
Checkpoint fold{k} must be evaluated only on source rows with
split_fold_{k}=test_fold_{k}. The five patient-disjoint test folds cover
13,694 examinations from 9,900 patients. They are not an ensemble; do not
apply one fold's model to the entire cohort or average the five models'
predictions for this benchmark.
Preserve the historical BU PNG preprocessing and the four-view order
LMLO -> LCC -> RMLO -> RCC, including its view/side metadata. A generic
Mayo-style DICOM loader is a different input pipeline and is not a drop-in
replacement for this benchmark.
The model returns raw, uncalibrated independent sigmoid probabilities for years 1 through 5. AUROC and AUPRC (average precision) are evaluated within each fold using horizon-specific follow-up masks and reported as unweighted five-fold means. C-index is reported as the mean and between-fold standard deviation. Do not replace these with pooled metrics, calibrated risks, or monotonic/cumulative post-processing.
| Horizon | Eligible examinations | Positive examinations |
|---|---|---|
| 1 year | 13,694 | 283 |
| 2 years | 12,724 | 495 |
| 3 years | 11,878 | 700 |
| 4 years | 11,113 | 888 |
| 5 years | 10,369 | 1,044 |
Model loading
The historical model state registers the same encoder under
image_encoder.* and aggregator.image_encoder.*. The Mammo-FM compatibility
loader verifies exact equality of these duplicate tensors before removing
only the duplicate namespace. The complete remaining state must then be
loaded strictly.
Download the compatible base checkpoint for architecture construction:
hf download batmanLab/Mammo-FM Mammo-FM_BatmanlabTrained_CLIP.tar \
--local-dir /path/to/checkpoints
The base checkpoint supplies the encoder architecture configuration. Loading the full risk checkpoint replaces the encoder and all risk-model tensors; the final predictions do not retain the base encoder's initialization. Optimizer and scheduler states are not used for inference. Load serialized training checkpoints only from a trusted source.
Validation status and use restrictions
Historical checkpoint identity and ported-code numerical parity are separate checks. Availability of these weights does not establish that the port has reproduced the historical BU held-out predictions or metrics. The fold-matched comparison must complete before such a claim is made.
The repository's model-weight license applies. The weights are for non-commercial academic research only and must not be used for clinical decisions.