A Graph Neural Network with Multi-head Attention for Universal Brain Disease Diagnosis from fMRI Images

Hyung-Jun Moon, Tae-Hoon Kang, Sung-Bae Cho

Abstract: Developing an almighty model for brain disease diagnosis is challenging due to the biological brain changes that should be detected in the brain across a wide range of conditions. Most models for brain disease diagnosis depend on the unique characteristics of each disease, which limits their applicability to universal brain diseases. In this paper, we propose a method that extracts all brain activities with functional connectivity (FC) graphs based on medical evidence and classifies them via a graph neural network (GNN) with multi-head attention for diagnosis. We generate FC graphs that contain both local and global features of brain changes, as well as short- and long-term dependencies of each disease. The FC graphs are then gone through the GNN to recognize correlations among various brain regions. Experiments on four benchmark disease datasets of autism, Alzheimer’s disease, ADHD, and schizophrenia show that the proposed method produces an average improvement of 1.6%p and up to 3%p in accuracy compared to the state-of-the-art methods. The results are verified with the medical knowledge on each disease, highlighting the effectiveness of the proposed method in advancing neurological research for brain disease diagnosis.

External IDs:dblp:conf/hais/MoonKC24