Dynamic functional connectivity analysis with temporal convolutional network for attention deficit/hyperactivity disorder identification

Abstract

Introduction: Dynamic functional connectivity (dFC), which can capture the abnormality of brain activity over time in resting-state functional magnetic resonance imaging (rs-fMRI) data, has a natural advantage in revealing the abnormal mechanism of brain activity in patients with Attention Deficit/Hyperactivity Disorder (ADHD). Several deep learning methods have been proposed to learn dynamic changes from rs-fMRI for FC analysis, and achieved superior performance than those using static FC. However, most existing methods only consider dependencies of two adjacent timestamps, which is limited when the change is related to the course of many timestamps.

Methods: In this paper, we propose a novel Temporal Dependence neural Network (TDNet) for FC representation learning and temporal-dependence relationship tracking from rs-fMRI time series for automated ADHD identification. Specifically, we first partition rs-fMRI time series into a sequence of consecutive and non-overlapping segments. For each segment, we design an FC generation module to learn more discriminative representations to construct dynamic FCs. Then, we employ the Temporal Convolutional Network (TCN) to efficiently capture long-range temporal patterns with dilated convolutions, followed by three fully connected layers for disease prediction.

Results: As the results, we found that considering the dynamic characteristics of rs-fMRI time series data is beneficial to obtain better diagnostic performance. In addition, dynamic FC networks generated in a data-driven manner are more informative than those constructed by Pearson correlation coefficients.

Discussion: We validate the effectiveness of the proposed approach through extensive experiments on the public ADHD-200 database, and the results demonstrate the superiority of the proposed model over state-of-the-art methods in ADHD identification.

Keywords: attention deficit/hyperactivity disorder; dynamics characteristics; functional connectivity; temporal convolutional network; temporal dependence.

Figure 1

Architecture of TDNet, including three components: (A) partitioning rs-fMRI time series via non-overlapping sliding windows, (B) a FC generation module (with three cascaded convolutional layers and a bilinear operation layer) to construct functional connectivity within each time window, and (C) a temporal convolutional network to capture temporal dynamics across all the time windows. With the output of the temporal convolutional network, a convolution is used to obtain these probabilistic values to fuse these time windows, followed by three fully-connected layers and a softmax activation for final disease identification. The left in TCN block is a dilated causal convolution with dilation factors D = 1, 2, 4 and filter size k = 3. The right is a TCN residual block. A 1 × 1 convolution is added when residual input and output have different dimensions.

Figure 2

Group difference between the learned FC and the traditional “Stationary FC”. Here, p-values less than 0.05 between ADHD and NC groups are set to 0 (corresponding to the green parts in the figure). The term A^t (t = 1, 2, ⋯ , 5) corresponds to the group difference based on dynamic functional connectivities learned on t-th sliding window by the proposed FC generation module in TDNet.

Figure 3

Top 15 brain functional connectivity patterns identified by our TDNet method in ADHD vs. NC classification on the KKI site.

Figure 4

Results of the proposed TDNet method with respect to different lengths of sliding windows in ADHD vs. NC on different sites.

Figure 5

Results of the proposed TDNet method with respect to different partitions of sliding windows in ADHD vs. NC on different sites.