pfd-gste 0.1.0 - Initial Python Package Release

Initial Python Package Release

This release corresponds to pfd-gste version 0.1.0, the first installable release of the reusable PFD–GSTE guidance components developed within the brain-tumour MRI research project.

Installation

pip install pfd-gste

Included components

• PathologyFocusedGate
• FeatureTokenGuidance
• PatchEmbed2d
• PatchTokenGuidance
• PFDGSTEVariantA
• PFDGSTEVariantB
• enable_mc_dropout
• mc_dropout_predict

Validation

• all nine package tests pass
• the wheel and source distribution pass strict metadata validation
• the package was validated through TestPyPI
• the production PyPI package installs successfully in a clean Python 3.12 environment
• both PFD–GSTE variants execute successfully after installation

Scope

This package contains only the reusable PFD–GSTE PyTorch components. The complete research repository remains the source for preprocessing, model training, held-out evaluation, explainability, trained checkpoints, experimental results and the local Flask prototype.

Research-use notice

This software is provided for research and educational use only. It is not a certified medical device and must not be used for clinical diagnosis, patient management or treatment decisions.

v1.0.0 — Initial Research Release

Initial Research Release

This is the first versioned research release of Mitigating Shortcut Learning in Brain Tumour MRI Classification, a project investigating pathology-guided hybrid CNN–Transformer models for four-class brain tumour MRI classification.

Included in this release

• Leakage-aware SHA1 deduplication, tight-crop preprocessing, and fixed train, validation, and test split generation
• Hybrid A with PFD-A and GSTE-A
• Hybrid B with PFD-B and GSTE-B
• Matched ablation without PFD-A and GSTE-A
• Matched ablation without PFD-B and GSTE-B
• Training, validation, and held-out test evaluation workflows
• Grad-CAM++ and attention-rollout explainability
• MC-dropout uncertainty support
• Recorded experimental results and misclassification analyses
• Trained model checkpoints managed through Git LFS
• A local Flask prototype for qualitative inspection
• Reusable PFD–GSTE guidance modules in pfd_gste/
• Research documentation, citation metadata, and an MIT licence

Reproduction

To reproduce the complete pipeline, download the dataset separately and follow the preprocessing, training, evaluation, and explainability instructions in the repository README.

Git LFS checkpoint notice

The trained .pt checkpoints are managed through Git LFS. Clone the repository using Git and retrieve the complete checkpoint files with:

git clone https://github.com/AnnyaB/HybridResNet50V2-RViT.git
cd HybridResNet50V2-RViT
git lfs pull

GitHub-generated ZIP or tarball archives may contain Git LFS pointer files rather than the complete checkpoint files.

Research-use notice

This software is provided for research and educational use only. It is not a certified medical device and must not be used for clinical diagnosis, patient management, or treatment decisions.