SPDGCN-KD: Knowledge Distillation-Based SPD Graph Convolutional Network for ADHD Diagnosis with Imbalanced fMRI Data

Attention deficit hyperactivity disorder (ADHD) is a common neurodevelopmental disorder. In recent years, graph neural networks on fMRI-based functional connectivities (FCs) have achieved great success in ADHD diagnosis. This is made by first constructing a population graph and then classifying the nodes on the graph to diagnose the states of subjects. Class imbalance is a prevalent issue in medical datasets, referring to a skewed distribution of classes. When the training data appear to be class imbalanced, the learned model is prone to be biased towards the major class, while overlooking the minor class which often represents patients that require greater attentions. However, existing methods used to resolve the graph class imbalance problem usually sacrifice the accuracy of the major class in order to improve the accuracy of the minor class. Therefore, for class-imbalanced node classification on graphs, we propose a SPD graph convolutional network based knowledge distillation (SPDGCN-KD) framework for ADHD diagnosis with FCs in this paper. Basically, we construct a population graph in the Riemannian manifold space and combine graph convolutional networks with the class re-balancing technique (i.e., balanced knowledge distillation) to mitigate the decline of the accuracy of the majority-class samples. The experimental results on the public ADHD-200 demonstrate the superiority of our SPDGCN-KD over several other competing methods for ADHD classification.