Four-grade lumbar foraminal stenosis classification on routine sagittal T2 MRI
Research model collection for Four-grade lumbar foraminal stenosis classification on routine sagittal T2 MRI: a patient-disjoint evaluation of grade-value auxiliary supervision and probability fusion. Author: Wuke Peng, ORCID 0009-0003-5791-3783.
Matching code distribution: v1.1.5. It retains all 81 Python/archived Python-text files from v1.1.4, including the same runtime/statistics program maps, frozen kernels/plans, fusion arithmetic and descriptive reporter. Earlier inference-compatibility verification was recorded for immutable code v1.1.2 at https://github.com/pengwuke/lss-grade-value-fusion/tree/ec9337ddb9ba5a0cbbd8f488e808e7e09a100a51; it is historical verification, not the latest documentation version. Historical compatible code v1.0.1 remains at https://github.com/pengwuke/lss-grade-value-fusion/tree/7dcb02165a332ddf6831b285598cd3000bbb8f6c. CODE_MODEL_COMPATIBILITY.json preserves those program and usage-document fingerprints and separately identifies later reporting/documentation changes. Program maps, rather than equality of whole documentation files, identify runtime compatibility.
The separate complete-fold reporting utility was added in code v1.1.3 and is unchanged in code release v1.1.5. Inference compatibility is determined by the retained runtime/model maps and does not require the new descriptive reporter.
Release scope and licence
Model weights retain release v1.0.0. This documentation correction is dated 2026-11-15 over the already public immutable documentation parent 1f81a4d04362f088628efd9520f9ba5e409cf276 (original weight release e780d57ac5280f025009188e4920a18bcd6626d7). It removes duplicated input-representation/geometry and artifact-citation paragraphs while retaining their content; weight files, models_manifest.json, fusion_config.json and the author-contribution grant remain unchanged. The author-approved standard CC BY-NC 4.0 applies only to the author's fine-tuned/trained contributions, with unchanged upstream rights; see LICENSE_AUTHOR_CONTRIBUTIONS.md. No raw MRI, ROI image/cache, patient/ROI prediction/probability row, patient-to-fold assignment table, optimizer state or original initial pretrained file is included.
Dataset v4 is cited at https://data.mendeley.com/datasets/rgb77xm3jf/4 (DOI 10.17632/rgb77xm3jf.4). As of 2026-11-09 its licence field says CC BY 4.0 while its description mentions noncommercial scientific use. Both observations are retained. This release does not represent unrestricted commercial permission or purport to relicense that dataset. The difference is not treated as a categorical prohibition on sharing noncommercial research models, and a dataset-author reply is not an automatic prerequisite for this research distribution.
Model identities
models_manifest.json lists 50 role files and 47 distinct tensor states: ten common5 core roles, twenty matched auxiliary-control roles, five common5 Phadke image-reference reconstruction roles, and fifteen source94 roles. Repeated roles are not independent trained models or cohorts. R10/R11 are frozen probability-fusion configurations, not additional checkpoints. All fifty weight bytes, model-manifest bytes and fusion_config.json are unchanged from internal R16. Forty source-checkpoint hashes had prior historical freeze records; ten ResNet/CORN source hashes were measured at preparation, not retroactively declared historically frozen.
Exports contain tensor state_dicts only. Load with weights_only=True and strict state loading into the matching architecture. The current matching code distribution v1.1.5 preserves the previously verified functional runtime/statistics programs, frozen kernels/plans and fusion arithmetic; earlier verification records retain their original scope. CODE_MODEL_COMPATIBILITY.json records the matching public code program hashes and preserves earlier fingerprints as historical references. The loader checks the model-side manifest/fusion fingerprints, each checkpoint size/hash, and matching architecture; it does not require code/model compatibility JSON files to be byte-identical.
Architectures, inputs and protocols
- CORAL CNN: author's frozen CNN definition, DICOM-derived tight IQ8, NoHE/G1, one channel at 64x64, normalisation 0.5/0.5.
- DeiT: timm deit_tiny_patch16_224.fb_in1k, four outputs, historical PNG/HE, three channels at 224x224, source float32 LUT normalisation order. The official upstream model card marks Apache-2.0 and ImageNet-1k pretraining. The original initial pretrained-file exact identity remains a historical gap.
- P4.1-C/SR1 image reconstruction: exact frozen timm ResNet-18 constructor with average pooling, dropout 0.3 and a three-logit CORN head. Source40 PNG crop uses a 40% box-relative margin on each side, PIL bilinear 224x224 and (x/255-0.449)/0.226 normalisation. CORN class scores retain the source 1e-6 floor without renormalisation. Historical timm resnet18 pretrained initial-file identity is not established by today's default model card.
Architectures are constructed with pretrained=False before loading complete fine-tuned states, so loading requires no new initial checkpoint download. This does not erase training-time upstream provenance. The common5 role must be used on its assigned held-out fold; averaging common5 models on their training cohort is not out-of-fold evaluation. source94 and SR1 have separate average-before-argmax rules. Callers must align ROI identity, representations and fold membership. Format/hash checks cannot establish them.
No reference grade is needed for image-model application. The ROI and its required representations are inputs; the predicted grade and four class scores are outputs. Reference grades are required only for performance evaluation. USE_AND_REPORTING_FAQ.md distinguishes these contracts.
Output semantics
The four-grade outputs support ROI classification and the frozen fusion rules; they have not been validated as calibrated clinical risk estimates. The training-only grade-value scalar is excluded from inference and does not predict longitudinal progression or treatment need.
INPUT_REPRESENTATION_LIMITS.md explains that the CNN helper expects a caller-prepared tight MONOCHROME2 ROI with the declared linear rescale contract, not a general DICOM decoder/display converter. G1 preserves the supplied ROI array height-to-width ratio, up to integer-size rounding; it has no spacing/orientation input and does not guarantee physical millimetre geometry or arbitrary anatomy preservation. Header, photometric and export equivalence are not certified by array/PNG checks. No whole-cohort header revalidation is claimed here.
CITATION.md distinguishes the unchanged v1.0.0 weight artifact from the code distribution and the actual immutable documentation revision. No manuscript DOI, acceptance or archival identifier is invented.
Historical checks and limits
The retained R4 record covers a fresh isolated Linux Python 3.11 CPU dependency installation on the same host, fifty strict role loads and fifty artificial-input forward interfaces. It is not a new machine, real-patient evaluation or historical training reproduction. R5 retained strict-load/bound-metadata checks. R6/R7 artificial common5 examples/path checks cover the ten common5 core roles. R9 records thirty runtime tests, seventy frozen numeric-rule comparisons and ten artificial-image CLI cases for the common5 float64 correction. Earlier records retain their actual origin and scope; the old internal candidate's unpublished/licence-pending fields are historical.
The prior documentation revision imported a separately locked descriptive reporting audit of existing frozen predictions, including each outer fold, fold mean/sample SD, pooled OOF and probability-quality summaries. This revision inherits those numerical files without recalculating them and only checks changed documentation/metadata. No model forward call, numerical aggregation, training, bootstrap interval, significance test, weight selection or clinical validation is added. Pure tensor export and successful architecture loading are not image-inference parity or external validation. Historical training/dependency/augmentation locking gaps remain disclosed.
The task is grading expert-selected sagittal T2 foraminal ROIs, not automatic localisation or a complete examination workflow. The study uses one cohort of 468 patients/2,978 ROIs after exclusions. Model-specific train/validation/outer roles were separated, but shared-grid development used information across cyclic folds; long same-cohort development and one seed therefore remain limits. Current archived results must not be presented as new fully development-independent confirmatory performance.
TASK_BASELINE_EVIDENCE.md separates task value, the rationale and remaining mismatches of the Phadke image-method reconstruction, whole-system comparisons, fixed-alpha controls, fusion trade-offs and descriptive complementarity. The original image-plus-morphometry system and original published test score are not reconstructed by those common-five-fold comparisons. QWK/MAE superiority over simpler natural-CORAL same-DeiT selected fusion remains unestablished, rather than equivalence or inferiority. The ten-outcome sensitivity applies only to that comparison, not to all Phadke comparisons or paper-wide inference; the 94-person source-protocol group is from the same source cohort, not new external validation. SR1 development-fold predictions from selected checkpoints were reused for fitting the source decoding rule; they had already served checkpoint selection and are not independent outer-test OOF predictions, regardless of the retained source variable names.
FIVE_FOLD_REPORTING.md and the accompanying aggregate evidence distinguish all five outer folds, arithmetic mean with sample SD and pooled OOF, and retain per-grade P/R/F1, normal/mild errors, moderate recall and severe underestimation. Original branch OOF probabilities were recovered; fixed and selected fusion probabilities were algebraically reconstructed with the already-frozen weights. All 19 resulting systems were aligned to the frozen decisions for log-loss, Brier and ECE10 descriptions. This does not mean 19 independent native probability files or new model inference. These are probability-quality and top-label calibration-error summaries, not evidence of classwise or clinically calibrated risk, successful recalibration or safety. Lower log loss or Brier alone does not isolate calibration; ECE10 depends on its confidence-bin definition. The described normalization/fusion is not a population-prior correction or recalibration. Training class balancing and scoring under the observed outer-fold grade distribution must not be conflated. Historical single-inner-VAL accuracy must not substitute for full OOF. READER_NAME_MAPPING.json maps readable system names to unchanged model/code IDs; fixed-alpha matched ablations remain distinct from the selected-alpha complete configuration.
Performance by age, sex, scanner/vendor, field strength or clinical site was not evaluated in the reported analyses and remains unknown. Overall QWK does not establish performance or equality of performance in these subgroups. The interface checks and grade/fold summaries do not provide demographic or acquisition-subgroup validation, and performance outside the source cohort remains unknown.
These files do not establish clinical utility, diagnostic safety, reader-time savings or readiness for deployment. They are provided as-is under the author-contribution grant and applicable unchanged upstream terms. Do not represent this collection as a separately validated clinical device.
Source DICOM metadata and model-facing inputs
The upstream image-reference repository warns that the source DICOM archive can retain date quasi-identifiers such as PatientBirthDate and StudyDate: https://github.com/rohanpa27/LSS_foraminal_stenosis_imaging . A limited tag-presence audit confirmed that such fields can be nonempty; it was not a whole-cohort de-identification certification, and no date values or patient keys are published here. The source DICOM archive must therefore not be described as fully anonymized merely because cropped images are used.
The inspected CNN/DeiT input paths use ROI pixels and grading targets; patient/study keys serve fold assignment and paired analysis. Birth/study dates and names are not predictor features in those inspected paths. The source archive was accessible, so this does not claim that all indirect identifiers were absent or inaccessible. This release contains no source DICOM, MRI/ROI cache, patient-level prediction/probability rows or date values. These facts do not change the author's MIT code grant, CC BY-NC4.0 model-contribution grant or retained upstream terms, and do not constitute a new institutional ethics determination.
For enquiries about access to the privately retained frozen analysis inputs, contact the corresponding author at 202628001@cquet.edu.cn. This is an enquiry route, not unrestricted public availability or a guarantee that any request will automatically receive data. It does not turn the source snapshots and public aggregate reports into a complete portable historical replay.
AI assistance
ChatGPT and Codex (OpenAI) assisted script implementation/debugging, frozen-result aggregation and consistency checking, interpretation, drafting/revision, translation/typesetting and code/model documentation, including the complete-fold reporter. This was not language editing alone. Outputs derive from executed programs and frozen artefacts; generative AI did not generate or alter MRI images, reference labels or experimental observations. The author retains responsibility. USE_AND_REPORTING_FAQ.md gives the assistance and verification scope.
Files and integrity
README.md and MODEL_CARD.md contain the same current card. models_manifest.json maps every role to its size, export SHA256, tensor-state fingerprint and source-checkpoint fingerprint. PUBLIC_DISTRIBUTION_MANIFEST.json states the exact repository scope; SHA256SUMS covers every regular file except itself. The Hugging Face revision history identifies the exact documentation commit, while the original weight release and all weight identities remain unchanged. Upstream full licences and attribution are retained under third_party_licenses/ and THIRD_PARTY_NOTICES.md. No private preparation credentials or patient-level rows are part of this repository.