Functional Connectivity Network Augmentation With Patch-Entirety Self-Attention GAN for ADHD Identification

Shengbing Pei, Wencong Jiang, Shuai Cao, Cunhang Fan, Wen Liang, Zhao Lv, Jihong Guan, Chao Zhang

Abstract: Functional connectivity network (FCN) describes the cooperative state of the brain and is often used to identify neurological diseases. However, the performance of FCN-based recognition models is limited by a small sample size for the high cost of patient data acquisition and labeling. Generative adversarial networks (GANs) can generate diverse samples to alleviate this problem. Due to the symmetry of FCN matrix, existing GAN-based data augmentation methods mainly adopt fully connected layers to learn flattened FCN data distribution, neglecting the functional collaboration structure within FCN. Here, we propose a novel patch-entirety self-attention GAN (PEGAN) for FCN data augmentation, thereby helping to improve brain disorder identification. PEGAN is a fully convolutional structure and contains a two-stage attention mechanism to learn the collaboration of brain regions. At the first stage, the FCN is partitioned into multiple patches, each patch is encoded by itself with self-attention, and this is to learn local consistency characteristic within a specific brain subfunction. At the second stage, the global self-attention involving all functional connections is performed, which is to learn the global collaborative characteristic of the entire brain. Meanwhile, a symmetry constraint is designed in the optimization objective of GAN to overcome the issue that the convolutional GAN model is not applicable to FCN data. With sufficient augmentation data, a dual-stream network is trained for classification. Experimental results on ADHD-200 dataset indicates that: 1) the proposed identification method with PEGAN demonstrates its superiority by comparing with existing ADHD identification methods; 2) the improved optimization objective function effectively solves the problem that FCN data are not suitable for convolution processing; 3) the reliability of the generated FCN data with PEGAN is verified from both qualitative and quantitative perspectives; and 4) the augmented FCN data can reduce noise interference in real samples and make the distribution of training samples more stable, which is the reason for improving the identification of ADHD patients.

External IDs:dblp:journals/tim/PeiJCFLLGZ25