The ABIDE (Autism Brain Imaging Data Exchange) benchmark challenges AI to classify Autism Spectrum Disorder (ASD) solely from resting-state fMRI brain scans.
ABIDE isn't just a folder of JPEGs. It's high-dimensional 4D functional imaging data. Understanding the pipeline is critical for building models.
Step 1: ParcellationWe map the 4D brain volume to an atlas (like Harvard-Oxford or AAL) to segment the brain into distinct functional regions.
Step 2: Time SeriesFor each ROI, we extract the mean BOLD signal over time. This raw signal (shown above) represents neural activity fluctuations.
Step 3: Functional GraphWe compute the correlation between every pair of regions. This matrix captures the "functional wiring" of the brain—the key to diagnosing ASD.
Comparing model performance on the standardized ABIDE I & II datasets. Note that the top result (Plymouth DL) was evaluated on a subset of participants.
| Model Rank | ABIDE I Acc | ABIDE I AUC | ABIDE II Acc | ABIDE II AUC | Modality | Eval Strategy |
|---|---|---|---|---|---|---|
#1 MAACNN Research | 75.1% | - | 72.9% | - | fMRICNN | k-fold |
#2 Plymouth DL Model Research | 98.0% | - | - | - | fMRIDeep Learning with XAI | k-fold |
#3 MCBERT Research | 93.4% | - | - | - | fMRI + PhenotypicMulti-modal CNN-BERT | LOSO |
#4 AE-FCN Research | 85.0% | - | - | - | fMRI + sMRIAutoencoder + Fully Connected Network | 10-fold |
#5 Multi-Atlas DNN Research | 78.1% | - | - | - | fMRIDeep Neural Network | k-fold |
#6 ASD-SWNet Research | 76.5% | 81.00 | - | - | fMRIShared-weight CNN | 10-fold |
#7 AL-Negat Research | 74.7% | - | - | - | fMRIGraph Neural Network | k-fold |
#8 BrainGNN Research | 73.3% | - | - | - | fMRIGraph Neural Network | 10-fold |
#9 GCN Research | 72.2% | 78.00 | - | - | fMRIGraph Convolutional Network | k-fold |
#10 Multi-Task Transformer Research | 72.0% | - | - | - | fMRITransformer | k-fold |
#11 PHGCL-DDGFormer Research | 70.9% | - | - | - | fMRIGraph Transformer | k-fold |
#12 SVM with Connectivity Features Research | 70.1% | 77.00 | - | - | fMRISupport Vector Machine | 10-fold |
#13 Deep Learning (Heinsfeld) Research | 70.0% | - | - | - | fMRIDeep Neural Network | k-fold |
#14 MVS-GCN Research | 69.4% | 69.01 | - | - | fMRIMulti-view Site Graph Convolutional Network | LOSO |
#15 Abraham Connectomes Research | 67.0% | - | - | - | fMRIConnectome Analysis | k-fold |
#16 Random Forest Baseline | 63.0% | - | - | - | fMRIEnsemble (Trees) | k-fold |
#17 BrainGT Research | - | 78.70 | - | - | fMRIGraph Transformer | 10-fold |
#18 DeepASD Research | - | - | - | 93.00 | fMRI + SNPsAdversary-regularized GNN | 10-fold |
The brain is fundamentally a network. Traditional CNNs struggle because functionally connected regions (like the default mode network) can be spatially distant.
GNNs excel by modeling ROIs as nodes and connectivity as edges. Recent models like BrainGT (Graph Transformer) and DeepASD (Adversarial GNN) capture subtle topological disruptions in ASD brains, such as hypo-connectivity in social networks, outperforming standard methods.
Comparing results on ABIDE is notoriously difficult due to three factors:
High accuracy isn't enough for clinical adoption; doctors need to know why. Recent SOTA models like the Plymouth DL Model (2025) use Explainable AI (XAI) techniques like gradient-weighted class activation mapping (Grad-CAM) to visualize decision-making.
Interestingly, these models highlight the visual processing regions (e.g., calcarine sulcus, cuneus) as critical discriminators, suggesting that sensory processing differences may be as fundamental to ASD as social cognition deficits.
1,112 resting-state fMRI datasets from 539 individuals with autism spectrum disorder (ASD) and 573 typically developing controls across 17 international sites. Multi-site neuroimaging data for autism classification and biomarker discovery.
1,114 datasets from 521 individuals with autism spectrum disorder (ASD) and 593 typically developing controls across 19 sites. Second large-scale release complementing ABIDE I with additional multi-site neuroimaging data.
Have you achieved better results on ABIDE? Or benchmarked a new GNN architecture? Help the community by sharing your verified results.