brainnetcnn

BrainNetCNN Model Skill

Overview

BrainNetCNN applies convolutional operators designed for adjacency matrices: edge-to-edge (E2E), edge-to-node (E2N), and node-to-graph (N2G). Use it when each subject is represented by a dense, consistently ordered ROI connectivity matrix and the target is categorical or continuous.

• Paper: Kawahara et al., 2017, BrainNetCNN: Convolutional neural networks for brain networks; towards predicting neurodevelopment
• NeuroClaw implementation: models/brainnetcnn/
• Input: dense FC matrix [subjects, ROI, ROI]
• Tasks: classification and regression
• Data adapter: shared with BNT

Research use only.

---

Input Contract

Prepare one file per subject:

data/braingnn_input/<atlas>/sub-<subject_id>.pt

Each file must contain:

{
    "subject_id": str,
    "atlas": str,
    "fc_matrix": Tensor[n_roi, n_roi],  # Fisher-z values
    "node_features": Tensor[n_roi, n_roi],  # accepted fallback
}

The shared BNT adapter applies tanh to recover Pearson correlations and zeros the diagonal. All subjects in one run must use the same atlas, ROI ordering, and matrix size.

Labels use CSV format:

subject_id,label
100001,0
100002,1

Change the columns with --subject-col and --label-col.

---

Quick Start

Validate data loading

python skills/brainnetcnn/scripts/train_reference.py \
  --atlas schaefer_100_7net \
  --labels-csv data/hcp_gender_labels.csv \
  --dry-run

Classification

python skills/brainnetcnn/scripts/train_reference.py \
  --atlas schaefer_100_7net \
  --labels-csv data/hcp_gender_labels.csv \
  --task classification \
  --nclass 2 \
  --fold 0 \
  --kfold 5 \
  --n-epochs 100 \
  --batch-size 16 \
  --device cuda

Regression

python skills/brainnetcnn/scripts/train_reference.py \
  --atlas aal_116 \
  --labels-csv data/hcp_age_labels.csv \
  --label-col age \
  --task regression \
  --fold 0 \
  --kfold 5 \
  --n-epochs 100 \
  --batch-size 16 \
  --device cuda