Official code and aggregate results for Cross-Patient Explanation Transportability in Medical Imaging.
TransXAI is an explainer-agnostic, post-hoc support-refinement method. It operates on frozen classifiers and frozen explanation maps: no classifier retraining and no shared explainer parameters are introduced during refinement.
At the primary exact 10% spatial budget, TransXAI increased the equal-cell mean Cross-Patient Transportability Score (CPTS) on both evaluated backbones:
| Backbone | Baseline CPTS | Refined CPTS | Delta CPTS | 95% CI |
|---|---|---|---|---|
| ResNet-18 | 0.4762 | 0.6012 | +0.1250 | [0.1214, 0.1285] |
| RegNet-X-400MF | 0.4857 | 0.5974 | +0.1117 | [0.1085, 0.1148] |
All 360 dataset-fold-explainer-backbone-budget cell means were positive. Exact mask cardinality and the frozen local-effect constraint had zero violations in the final paired panel.
- Datasets: SIIM-ACR, ISIC 2016, and PAD-UFES-20.
- Backbones: ResNet-18 and RegNet-X-400MF.
- Explainers: Grad-CAM, LayerCAM, Integrated Gradients, LRP, RISE, and Extremal Perturbation.
- Five grouped outer folds and 128 out-of-fold donors per dataset and fold.
- Exact spatial budgets: 10% (primary) and 20% (prespecified robustness analysis).
- Matching representation: the second convolutional stage, with spatial size 40 x 40 for 320 x 320 inputs.
- Intervention representation: the third convolutional stage, with spatial size 20 x 20.
- Recipient sets: 12 support, 12 validation, and 10 untouched query recipients.
- Frozen seed:
20260906.
The frozen implementation identifier T-CPT remains in some configuration keys, output paths, and historical manifests. It refers to the method named TransXAI in the paper.
configs/ frozen experiment contracts
docs/ data, implementation, and reproduction notes
results/figures/ publication figures generated from aggregate results
results/manifests/ immutable run manifests and artifact hashes
results/tables/ aggregate and cell-level paper tables
scripts/ executable single-cell Colab scripts
tests/ release-integrity checks
Raw clinical images, model checkpoints, HDF5 explanation maps, local caches, and patient/donor-level records are intentionally not distributed.
Each file in scripts/ is a self-contained Python cell intended for Google Colab. The default project root is:
/content/gdrive/MyDrive/Colab Notebooks/CPETChange PROJECT_ROOT consistently if a different Drive location is used.
| Stage | Script | Hardware | Purpose |
|---|---|---|---|
| 1 | 01_colab_setup.py |
CPU | Public environment and project-layout setup |
| 2 | 02_preprocessing_and_splits.py |
CPU | Acquisition, auditing, and grouped folds |
| 2B | 02B_repair_pad_ufes20.py |
CPU | Frozen PAD-UFES-20 distribution repair used in the study |
| 3 | 03_train_resnet18.py |
GPU | Five-fold ResNet-18 classifiers |
| 3R | 03R_train_regnet_x_400mf.py |
GPU | Five-fold RegNet-X-400MF classifiers |
| 4 | 04_explainers_resnet18.py |
GPU | ResNet-18 OOF explanation maps |
| 4R | 04R_explainers_regnet_x_400mf.py |
GPU | RegNet OOF explanation maps |
| 6B | 06B_transxai_resnet18.py |
GPU | ResNet TransXAI refinement at 10% and 20% |
| 6R | 06R_transxai_regnet_x_400mf_budget10.py |
GPU | RegNet refinement at 10% |
| 6R | 06R_transxai_regnet_x_400mf_budget20.py |
GPU | RegNet refinement at 20% |
| 7 | 07_confirmatory_statistics_resnet18.py |
CPU | ResNet confirmatory analysis |
| 7R | 07R_backbone_replication_statistics.py |
CPU | Paired architectural replication |
| 8 | 08_reviewer_controls.py |
GPU | Compact controls and ablations |
The initial ResNet training script also processes BUS-UCLM as a development-only smoke dataset. BUS-UCLM is excluded from the confirmatory TransXAI pipeline and from all paper results.
For a syntax and release-integrity check:
python tests/validate_release.pyThe aggregate tables and publication figures are already included under results/. Full recomputation requires the external datasets and the large intermediate artifacts generated by stages 3-6. Statistical stages 7 and 7R additionally require the frozen donor-level Parquet outputs, which are withheld because they contain sample and grouping identifiers.
See docs/REPRODUCIBILITY.md for the full contract and docs/IMPLEMENTATION.md for the exact implemented estimand.
No medical image is redistributed. Users must obtain each dataset from its original source and comply with the corresponding license and terms. See docs/DATASETS.md.
If you use this repository, cite the accompanying paper. Machine-readable metadata are available in CITATION.cff.
The code is released under the MIT License. Dataset licenses and third-party model or explainer licenses remain with their respective owners.