Brain Tumour Pfd Gste Hybrid A

--- license: mit library_name: pytorch pipeline_tag: image-classification tags: - medical-imaging - brain-mri - brain-tumour-classification - pytorch - cnn-transformer - explainable-ai - gradcam - attention-rollout - shortcut-learning - pfd-gste metrics: - accuracy - f1 --- # Brain Tumour PFD-GSTE Hybrid A This repository contains the trained **Hybrid A — PFD-A + GSTE-A** checkpoint from the project: **Mitigating Shortcut Learning in Brain Tumour MRI Classification** BSc Artificial Intelligence Project, University of Hertfordshire Author: Riya Basak Supervisor: Dr Kheng Lee Koay ## Links - Main GitHub repository: https://github.com/AnnyaB/HybridResNet50V2-RViT - Live Hugging Face Space: https://huggingface.co/spaces/AnnyaaB/brain-tumour-pfd-gste-demo - This model repo: https://huggingface.co/AnnyaaB/brain-tumour-pfd-gste-hybrid-a ## Model Summary Hybrid A uses PFD-A and GSTE-A, where pathology-focused guidance gates the transformer-token pathway and reweights 49 CNN feature tokens. The checkpoint is a PyTorch hybrid CNN–Transformer model for four-class brain MRI classification: - glioma - meningioma - pituitary - notumor ## Files - `best_model.pt` — trained PyTorch checkpoint. - `architecture.py` — model architecture source file. - `xai.py` — model-specific XAI helper file. - `models/hybrid_model.py` — loading wrapper used by the app workflow. - `pfd_gste/` — local PFD-GSTE guidance modules used for release reproducibility. - `model_config.json` — model metadata and preprocessing configuration. - `SHA256SUMS` — checkpoint checksum for verification. ## Evaluation Summary Recorded held-out test-set performance from the project repository: | Model | Test Accuracy | Macro F1 | |---|---:|---:| | Hybrid A — PFD-A + GSTE-A | 0.9875 | 0.9875 | ## Intended Use This model is released for research reproducibility, educational inspection, and comparison with the associated guided and ablation variants. ## Limitations - The model was evaluated on a single benchmark curation. - The checkpoint is not externally clinically validated. - Grad-CAM++ and attention rollout are qualitative inspection tools, not clinical annotations. - Outputs may be incorrect and should not be used for medical decision-making. - The dataset is not redistributed in this model repository; users should follow the GitHub reproduction instructions to obtain the benchmark dataset s...