What problem does it solve? Running the BrainNetworkTransformer (BNT) model on neuroimaging data normally requires adapting the original hydra/PyG-based codebase to your own datasets. This Skill provides a NeuroClaw reimplementation that loads preprocessed ROI-level FC matrices directly, trains BNT with DEC pooling for classification or regression, and evaluates with K-fold cross-validation. ## Core Features & Use Cases - Dense FC input without PyG: Loads NeuroClaw .pt files (Fisher-z transformed FC matrices), inverts to Pearson r via tanh, and feeds dense [B, N, N] tensors to a pure PyTorch model. - Classification and regression: Supports phenotype classification (e.g., gender) and regression (e.g., age) with z-score label standardization and stratified K-fold CV. - Interpretability outputs: Exposes transformer attention weights and DEC soft cluster assignments for ROI importance and clustering analysis. - Optional T1 fusion: Concatenates z-scored gray matter volume as an extra node feature column when --include-t1 is set. - Use Case: Given HCP subjects preprocessed into data/braingnn_input/schaefer_100_7net/, train a 5-fold CV classifier predicting gender from FC matrices with T1 volume fusion. ## Quick Start Ask the agent to train the BNT model on the aal_116 atlas with the HCP gender labels CSV using 5-fold cross-validation and 50 epochs.
