What problem does it solve? Applying convolutional neural networks to brain connectivity data requires architectures designed for adjacency matrices rather than images. This Skill trains and evaluates BrainNetCNN, which uses edge-to-edge, edge-to-node, and node-to-graph convolutions, on dense ROI-by-ROI functional connectivity matrices for classification or regression tasks. ## Core Features & Use Cases - Connectome-Specific CNN: Applies E2E, E2N, and N2G convolution blocks to dense Fisher-z connectivity matrices with configurable channel sizes, dropout, and learning rate. - Classification and Regression: Supports both categorical targets (e.g., gender classification) and continuous targets (e.g., age regression) with subject-level k-fold cross-validation. - Standardized Input Contract: Consumes per-subject .pt files containing fc_matrix tensors and a CSV label file, with a shared BNT data adapter handling tanh transformation and diagonal zeroing. - Use Case: A researcher with HCP Schaefer-100 atlas connectivity matrices wants to predict subject gender; they run the reference trainer with 5-fold cross-validation on GPU and obtain per-fold checkpoints with best metrics. ## Quick Start Train BrainNetCNN on my Schaefer-100 connectivity matrices using the labels in data/hcp_gender_labels.csv with 5-fold cross-validation on GPU.
