ADHD classification via auto-encoding network with non-imaging data fusion

With an existing binary hypothesis testing frame-work, we propose a modified auto-encoding (AE) network for ADHD classification using both Functional Connectivity (FC) and non-imaging data. For the presented AE network, it learns high-level features from the training FC data through a traditional auto-encoder. Here, a classification subnetwork is added to supervise the high-level features with their corresponding labels, thereby making these features more discriminative. The non-imaging data, including age, gender and Intelligence Quotient (IQ), are also used in this subnetwork. It further increases the subnetwork's robustness. We generate two sets of high-level features under the binary label hypotheses of test data, respectively. Then, a variability score comparison is provided on the high-level feature sets between two hypotheses. A label prediction of the test data is finally performed in terms of the comparison result. On the test of ADHD-200 datasets, our method achieves a high average accuracy of 99.7% and outperforms several state-of-the-art ADHD classification methods.