brain-connect-gcn

Adversarial Brain Mode Network — site-invariant ASD classification from resting-state fMRI.

20-model LOSO ensemble trained on ABIDE I (1,102 subjects, 20 acquisition sites, 3 atlases). Achieves AUC 0.7298 on CC200 cross-site evaluation — every subject is test-only exactly once, trained on all other sites.

Live demo: BrainConnect-ASD Space
Checkpoints: stored in the Space repo under checkpoints/

---

Architecture

Input: BOLD time-series (196 TRs × N ROIs)
  ↓
Preprocessing: z-score → FC matrix (Pearson r → Fisher z) → adjacency thresholding
  ↓
Brain Mode Decomposition: K=32 learnable spectral modes (graph wavelet basis)
  ↓
GCN Encoder: 2-layer graph convolution, hidden_dim=128, dropout=0.3
  ↓
Gradient Reversal Layer (Ganin 2016): removes site-specific scanner artifacts
  ↓
ASD Classifier head → p(ASD) ∈ [0, 1]
  ↓
20-model ensemble mean → final prediction

Component	Detail
Total parameters	~105K
Learnable modes	K = 32
Hidden dim	128
Dropout	0.3
Atlases	CC200 (200 ROIs) · AAL (116 ROIs) · Harvard-Oxford (111 ROIs)
Loss	Cross-entropy + GRL adversarial site loss
Training	150 epochs · batch 32 · lr 5e-4 · AdamW

---

Why Adversarial Training?

ABIDE I was collected across 20 sites with different scanners, protocols, and demographics. A naive classifier learns site identity as a shortcut — it generalizes poorly to unseen institutions.

The Gradient Reversal Layer (GRL) flips the gradient sign for the site prediction head during backprop, forcing the encoder to produce representations that are maximally confusing to a site classifier. Scanner artifacts and acquisition-specific signals cannot leak into the ASD prediction.

This is validated empirically: the LOSO protocol uses a model trained on 19 sites to predict the 20th — it has never seen that institution's data, scanner, or demographic distribution.

---

Results

Metric	Value
CC200 LOSO AUC	0.7298
HO LOSO AUC	0.7212
AAL LOSO AUC	0.6959
Subjects	1,102 (ABIDE I)
Sites	20
Evaluation	Cross-site LOSO — every subject is test-only

vs. Published ABIDE Baselines

Model	AUC	Split
SVM + FC (Plitt 2015)	0.71	Same-site
BrainNetCNN (Kawahara 2017)	0.74	Same-site
GCN + FC (Ktena 2018)	0.70	Same-site
ABIDE site-specific SVM	0.76	Same-site
BrainConnect-ASD (ours)	0.7298	Cross-site LOSO

All prior baselines use same-site train/test splits — a fundamentally easier evaluation. Cross-site LOSO is the clinically relevant bar.

---

Training Hardware

Trained on AMD MI300X (192GB HBM3) via DigitalOcean, ROCm 7.0, PyTorch 2.5.1+rocm6.2.

All 20 LOSO folds across 3 atlases (60 total training runs) were executed in parallel on the MI300X — a workload that would take days on CPU completed in hours.

End-to-end inference (preprocessing + 20-model ensemble): ~20ms per subject on MI300X.

---

Usage

# Checkpoints are in the BrainConnect-ASD Space repo
from huggingface_hub import hf_hub_download
from brain_gcn.tasks import ClassificationTask

ckpt = hf_hub_download(
    repo_id="lablab-ai-amd-developer-hackathon/BrainConnect-ASD",
    filename="checkpoints/cc200/adv_brain_mode_k32_site_cc200_loso_nyu/best.ckpt",
    repo_type="space"
)
task = ClassificationTask.load_from_checkpoint(ckpt, map_location="cpu", strict=False)
task.model.eval()

# bw_t: (1, 30, 200) windowed BOLD · adj_t: (1, 200, 200) FC adjacency
import torch
with torch.no_grad():
    out = task.model(bw_t, adj_t)
    p_asd = torch.softmax(out, -1)[0, 1].item()

---

Citation

BrainConnect-ASD — AMD Developer Hackathon 2026
Raghav Aryen · lablab.ai · AMD MI300X
https://huggingface.co/spaces/lablab-ai-amd-developer-hackathon/BrainConnect-ASD