Uncertainty‐Aware Graph Self‐Training for Autism Spectrum Disorder Classification in Multiple Centers

ABSTRACT Classical self‐training methods for graph convolutional networks (GCNs) assume that both labeled and unlabeled data follow the identical distribution. However, these works do not work well when they are used in medical applications such as classifying autism spectrum disorder (ASD) in multiple centers, where the unlabeled samples from different imaging centers have varying distribution shifts from the labeled samples. To this end, we propose uncertainty‐aware graph self‐training (UA‐GST) by extending graph self‐training to a new situation, in which the labeled data come from one imaging center and the unlabeled data come from several other imaging centers. Specifically, an uncertainty‐aware mechanism is proposed to select unlabeled centers and adversarial domain adaptation is introduced into graph self‐training to reduce domain shift between centers. With the progression of self‐training, more pseudo‐labeled test samples are gradually included in the training set, and a final model is finally trained on all the labeled and pseudo‐labeled samples. Considering the over‐confidence issue of the classifier, an evidential GCN is further proposed to estimate the uncertainty of the pseudo‐labels using Dempster–Shafer (D–S) evidence theory. It is evaluated on the Autism Brain Imaging Data Exchange (ABIDE). Experimental results verified the effectiveness of the proposed method in classifying ASD in multiple centers, outperforming existing state‐of‐the‐art methods.