LSS Protocol V3 ordinal-grading checkpoints

This public model repository contains the 20 canonical Protocol V3 checkpoints used in the primary analyses of:

Ordinal grading of lumbar foraminal stenosis on sagittal MRI: patient-disjoint evaluation of CORAL refinement in a dual-branch model with review prioritisation.

Four model branches × five patient-disjoint folds:

reproduced_baseline_cnn
reproduced_baseline_deit
coral_cnn
coral_auxiliary_cnn

The three primary fused systems are reconstructed as follows:

Reproduced baseline: 0.60 × DeiT + 0.40 × baseline CNN
Ordinal model: 0.60 × DeiT + 0.40 × CORAL CNN
Final auxiliary model: 0.60 × DeiT + 0.40 × CORAL + auxiliary CNN

Locked release provenance

Release version: v1.0.0
Code commit used for model-release provenance: b63de909b592204e8490c32b078617e1b2baae4b
Source dataset DOI: https://doi.org/10.17632/rgb77xm3jf.4
Author: Wuke Peng
ORCID: 0009-0003-5791-3783

The source code repository remains private during peer review. It will be made public after the release and authorship checks are completed.

Files

Each public checkpoint contains CPU model_state_dict tensors and minimal release metadata only. The repository does not contain:

source MRI or DICOM files;
ROI images;
patient identifiers;
optimiser states;
local-machine paths;
clinical reports;
predictions used as patient-level clinical decisions.

See manifest.csv, manifest.json, and CHECKSUMS.sha256 for file-level provenance and integrity.

Evaluation setting

468 evaluable patients
2,978 expert-defined ROIs
fixed patient-disjoint five-fold out-of-fold evaluation
primary optimisation seed: 42
internal retrospective validation on one public dataset

Intended use

Research reproducibility, non-commercial scientific validation, and method comparison.

Limitations and clinical boundary

These checkpoints are not a clinical device and must not be used for autonomous diagnosis, clinical triage, treatment decisions, or exclusion from radiologist review. The models were evaluated on expert-defined ROIs from a single public dataset and have not undergone external, prospective, subgroup, or reader-performance validation.

Citation

A DOI should be generated from the Hugging Face repository settings only after the uploaded files and metadata have been checked. Once generated, cite the DOI shown on this model repository.

Repository identifier before DOI generation:

https://huggingface.co/wuke2024/lss-foraminal-ordinal-grading-models

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

wuke2024
/

lss-foraminal-ordinal-grading-models