jrieke/cnn-interpretability

Implements four gradient-based and occlusion-based visualization methods (sensitivity analysis, guided backprop, occlusion, area occlusion) as reusable PyTorch utilities that work on both 2D and 3D medical images to highlight clinically-relevant decision regions. The pipeline trains a 3D CNN on preprocessed ADNI MRI scans (registered to ICBM template via ANTs) and applies interpretability techniques to validate that learned features correspond to known Alzheimer's pathology markers like temporal gyrus atrophy. Code is organized around Jupyter notebooks with modular `.py` interpretation methods, enabling application to custom models beyond the included trained checkpoint.

178 stars. No commits in the last 6 months.