How lin-lcx/hgcn works
Project Type and Purpose
• This is a research‐oriented machine learning pipeline for cancer survival prediction.
• It integrates a Hybrid Graph Convolutional Network (HGCN) with an online masked autoencoder to handle multimodal patient data (pathological images, clinical records, genomic profiles).
High-Level Component Breakdown
A. Data Acquisition & Storage
• External Data Sources: cBioPortal (genomics), GDC (slides & clinical).
• Local Data Folder:
• data_split/ (precomputed train/test splits)
• Supplementary files (e.g., esca_split.pkl, etc.)
B. Data Preprocessing
• cut_and_pretrain/
• cut_and_pretrain.py: patches extraction & autoencoder pretraining
• utils/filters.py, utils/general.py: image filtering, augmentation, helper functions
• gendata.ipynb
• Notebook for merging modalities into unified graph input format
C. Model Definition & Utilities
• HGCN_code/
• mae_model.py: masked autoencoder and HGCN architecture
• mae_utils.py: masking strategy, reconstruction loss functions
• util.py: graph construction routines, feature normalization
D. Training & Evaluation
• train.py (entry point)
• Loads preprocessed graphs, applies data_split if enabled
• Instantiates model, optimizer, scheduler
• Runs training loop, validation, test inference
• Saves trained weights and prediction outputs
Relationships and Data Flow
• Raw data (images, genomics, clinical) → cut_and_pretrain → patch embeddings
• Patch embeddings + clinical + genomic features → gendata.ipynb → multimodal graph files
• Graph files + data_split → train.py → HGCN_code modules → model training
• Trained model ↔ online masked autoencoder paradigm for modality imputation at inference
• Final output: survival risk predictions written to disk
Architectural Patterns & Key Design Principles
• Modular Pipeline: clear separation of Data Acquisition → Preprocessing → Model → Training
• Graph-based Learning: uses PyTorch Geometric to build modality-specific graphs and hypergraphs
• Masked Autoencoding: integrated into the model for robust missing-modality handling
• Configurability: train.py parameters, if_fit_split flag, easily switch between full/missing modalities