These 14 notebooks document the study workflow and aggregate analyses. Every committed notebook has no outputs, execution counts, attachments, patient data, or machine-specific paths.
Internal preprocessing independently centers P, C1, C2, and C3 crops and performs no explicit interphase registration. The corrected primary clinical/fusion analysis retains unresolved sex as missing and uses fold-local medians. The historical non-male-to-zero rule is sensitivity-only.
Create ignored configs/paths.yaml from configs/paths.example.yaml only when user-obtained data
or authorized local artifacts are available. Patient-linked outputs must remain outside Git.
| Order | Notebook | Classification |
|---|---|---|
| 1 | Internal cohort and splits | Public dataset plus configured private storage |
| 2 | Independent-phase preprocessing and sex provenance | Public dataset plus configured private storage |
| 3 | W3 single-phase CNN | Historical workflow requiring private artifacts |
| 4 | W4 multiphase CNN | Historical workflow requiring private artifacts |
| 5 | W5 radiomics-LightGBM | Historical workflow requiring private artifacts |
| 6 | Late fusion and internal evaluation | Public aggregate reanalysis; private artifacts required for refitting |
| 7 | External data acquisition | Public acquisition into private storage |
| 8 | External DICOM-to-NIfTI | Public dataset plus configured private storage |
| 9 | External tumor-mask preprocessing | Public dataset plus operator-controlled private review |
| 10 | External mask-density QC | Historical workflow requiring private images and masks |
| 11 | External liver-mask adapter | Historical per-case method interface |
| 12 | External frozen inference | Public aggregate reanalysis; complete historical CNN inference unavailable |
| 13 | Radiomics imputation controls | Public aggregate reanalysis; private artifacts required for rerun |
| 14 | Explainability and aggregate figures | Public aggregate figure generation; private artifacts required for case-level explainability |
- W3 fold 1 and W4 fold 0 checkpoints were not retained.
- Source geometry, spacing, and phase-specific crop coordinates were not retained.
- Public C2 crop-retention findings apply only to C2.
- W5 is crop-restricted radiomics and its shape outputs are not verified physical-unit measures.
- Exact full fusion contains W4 and W5 probability triplets, age, and sex; W3 is excluded.
- External values are HCC sensitivities under model-defined unavailable-covariate scenarios.
See reproducibility and artifact contracts.