How dilant123/lowcost-ml-cancer-padufes20 works
Project type and purpose
• This is a research-grade, end-to-end multimodal machine-learning pipeline for classifying skin lesions (PAD-UFES-20).
• It ingests clinical metadata (CSV) and medical images, extracts features, applies preprocessing (SMOTE, PCA, feature selection), trains multiple models, builds a weighted ensemble, and outputs results and visualizations.
File structure
• Top-level artifacts:
• Notebooks (.ipynb) for exploration and post-article code
• config.yaml (configuration parameters)
• requirements.txt (dependencies)
• Directories: graficos/, plots_enhanced/ (visual outputs)
• From the README, the canonical project layout under PAD_UFES_Multimodal/ is:
• data/ (raw inputs: metadata.csv, imagenes/ folder)
• resultados/ (combined dataset and exploration stats)
• graficos/ (static plots)
• plots_enhanced/ (interactive plots)
• modelos/ (saved model files and metrics)
Key architecture insights from the README
• Single-process pipeline orchestrating:
a. Data ingestion (metadata + images)
b. Feature extraction (image processing + clinical features)
c. Class balancing (SMOTE)
d. Dimensionality reduction & feature selection (PCA, RFE, variance, mutual information)
e. Model training (XGBoost, Random Forest, Extra Trees, SVM)
f. Ensemble learning (custom weighted, stacking, voting)
g. Output generation (models, CSV/JSON results, PNG/HTML plots)
• Configuration and parameters centralized in config.py (paths, hyperparameters, test/val splits).