How benkeel/vae_lung_lesion_bmvc works
Project type and purpose
• This is a research-oriented machine-learning pipeline for variational autoencoders (VAEs) applied to lung lesion images, followed by downstream tasks (latent-space clustering, MLP classification).
• The README stresses image preprocessing, VAE training (Gaussian & Dirichlet variants), latent extraction, clustering, and hyperparameter searches for MLPs.
File structure
Top-level directories:
• Preprocessing – Jupyter notebooks for converting DICOM→NumPy, data splitting, ROI cropping, metadata files.
• VAE – Python training scripts for various VAE architectures and joint-loss classifiers, plus an Excel hyperparameter log.
• Clustering and MLP – Notebooks and scripts for latent-space exploration, clustering, and MLP hyperparameter searches.
• data_2classes.npy / data_3classes.npy, labels2.npy / labels3.npy, meta CSVs – serialized data artifacts.
• interesting_examples – output artifacts (e.g., GIFs of latent traversals).
Configuration/build artifacts: minimal; research code uses notebooks and standalone Python scripts.
Extract additional insights from the README
• Preprocessing stage generates ROI-cropped 64×64 numpy arrays and metadata splits (malignant vs benign/non-malignant).
• VAE stage trains multiple flavors: Gaussian, Dirichlet, joint VAE+classifier.
• Clustering stage performs PCA/t-SNE, K-Means/CLASSIX, and MLP performance grids.