Explainable AI Tool Improves Brain Disorder Screening from 4D fMRI

Functional magnetic resonance imaging (fMRI) captures changes in brain activity over time, but the resulting image sequences can be difficult to analyze. Screening for brain disorders requires assessing both where activity occurs and how it changes over time. Clinicians also need insight into the evidence behind an artificial intelligence (AI) prediction, not just its diagnostic label. Researchers have now developed an AI approach designed to combine these spatial and temporal features while making screening more transparent and efficient.

At InfoLab, Sungkyunkwan University (SKKU; Seoul, South Korea), researchers developed the 4D fMRI CrossFormer (4DfCF), a vision transformer architecture for four-dimensional fMRI data. The model analyzes spatial and temporal patterns of brain activity together. It examines activity at different scales and learns relationships between both nearby and distant brain regions. 

The researchers also applied an explainable AI technique to the model’s predictions. It produces visual maps showing which brain regions contributed most strongly to each prediction. These maps give clinicians and researchers information to review alongside the model’s output, without replacing a physician’s judgment. 

The team evaluated 4DfCF using three benchmark fMRI datasets containing scans from people with attention-deficit/hyperactivity disorder, Alzheimer’s disease or autism spectrum disorder. The model outperformed the comparison models assessed across all three datasets. It achieved an F1 score of 96.28% on the Alzheimer’s disease dataset identified as ADNI. 

A model initially trained on one fMRI dataset also learned another dataset faster and achieved improved performance. The main 4DfCF model has approximately 10.34 million parameters, while the smaller 4DfCF-T version has about 4.18 million. Both require fewer computational resources than several larger comparison models. The research was published in the IEEE Journal of Biomedical and Health Informatics. 

The findings come from benchmark datasets rather than clinical deployment. Validation across hospitals and patient populations remains necessary. The results also suggest that pretrained models might eventually be adapted to other brain disorders and neuroimaging tasks.

Related Links
Sungkyunkwan University