Standalone reproducibility package for the external paired-scanner validation of Paired-Acquisition Neural Factorization on the Multi-Scanner Canine Cutaneous Squamous Cell Carcinoma histopathology dataset.
I use this repository to test Paired-Acquisition Neural Factorization as a representation-identifiability method. The study asks whether a locked paired-acquisition neural factorization objective can transfer from SCORPION to an independent five-scanner benchmark and still separate tissue identity from acquisition provenance.
The claim is deliberately narrow: the method reduces scanner identifiability in the scanner-suppressed tissue factor while preserving same-region retrieval and cross-scanner tissue agreement. I do not claim that this proves disease biology, clinical validity, diagnostic equivalence, or complete factor separation.
Paired-Acquisition Neural Factorization reduced scanner identifiability on a locked five-fold external test while preserving same-region retrieval.
| Metric | Paired reference | Factorized dep20 | Difference |
|---|---|---|---|
| Scanner probe accuracy | 0.752868 | 0.361408 | -0.380347 sample-blocked contrast |
| Pair cosine average | 0.696022 | 0.729961 | +0.033256 sample-blocked contrast |
| Pair cosine worst | 0.627300 | 0.656736 | +0.033104 sample-blocked contrast |
| Retrieval top-1 average | 0.930637 | 0.933392 | +0.002326 sample-blocked contrast |
| Retrieval top-1 worst | 0.881242 | 0.884431 | +0.001731 sample-blocked contrast |
All predefined success criteria passed over 44 biological sample blocks.
The public canine SCC files used here are downsampled to approximately 4 µm per source pixel, rather than the original scanner sampling of roughly 0.22--0.26 µm per pixel. The benchmark therefore evaluates scanner-factor suppression and same-region retrieval at coarse tissue-context scale. It does not establish preservation of nuclear morphology, native 20×/40× robustness, or cellular-scale scanner invariance. Resizing for an encoder cannot restore detail removed before public release.
See the magnification-mismatch audit for the deterministic physical-field-of-view audit and the still-pending local sensitivity gate.
| Audit metric | Mean |
|---|---|
| Acquisition scanner probe | 0.865098 |
| Acquisition tissue retrieval | 0.180627 |
| Acquisition effective rank | 13.756 |
| Cross-covariance RMS | 0.089831 |
Interpretation: scanner identity remained strongly available in the acquisition-specific factor while same-region tissue retrieval was concentrated in the scanner-suppressed tissue factor. This supports a factor-separation interpretation rather than simple representational destruction.
- Dataset inspection and annotation correspondence scripts.
- Geometry qualification and P1000 orientation-normalization scripts.
- Patch extraction manifest generation.
- Frozen encoder analysis scripts.
- Locked validation and five-fold test runners.
- Compact result tables and JSON summaries.
The repository intentionally excludes raw TIFFs, extracted JPEG patches, NPZ feature archives, model checkpoints, and full run directories. These artifacts are large or regenerable from the public dataset and scripts.
- Download the public Multi-Scanner Canine SCC dataset.
- Build or verify the geometry-qualified patch manifests.
- Extract orientation-normalized patches locally.
- Extract DINOv2 frozen features.
- Run
experiments/external_multiscanner/run_canine_paired_acquisition_factorization_crossfold.py. - Run
scripts/external_multiscanner/analyze_canine_paired_acquisition_factorization_crossfold.py.
To reproduce the tracked physical-field-of-view audit without the source TIFFs:
python scripts/external_multiscanner/audit_canine_magnification_mismatch.py --checkSee docs/results.md for the frozen result statement and claim boundary.
https://github.com/matthewvaishnav/computational-pathology-research
This is a representation-identifiability and paired-acquisition validation study. It is research only and is not clinical, diagnostic, or patient-care software.