A Confounding Factors-Inhibition Adversarial Learning Framework for Multi-site fMRI Mental Disorder Identification

TL;DR

Multi-site rs-fMRI classification suffers from site-specific confounding due to scanning and population differences. The authors present MSalNET, an adversarial framework with three components: (i) Node Information Assembly (NIA) for FC feature learning, (ii) adaptive site feature extraction via autoencoder, and (iii) a GAN-like adversarial objective that balances disease classification with site regression using the loss $L_t = L_C + \alpha / L_R$. On ABIDE-I and ADHD-200, MSalNET achieves 75.56% and 68.92% accuracy, respectively, while reducing site variability and enabling ROI-level interpretability of discriminative biomarkers. The approach offers data-driven harmonization across sites and demonstrates strong generalization, with interpretable brain region contributions aligning with established neurobiological findings.

Abstract

In open data sets of functional magnetic resonance imaging (fMRI), the heterogeneity of the data is typically attributed to a combination of factors, including differences in scanning procedures, the presence of confounding effects, and population diversities between multiple sites. These factors contribute to the diminished effectiveness of representation learning, which in turn affects the overall efficacy of subsequent classification procedures. To address these limitations, we propose a novel multi-site adversarial learning network (MSalNET) for fMRI-based mental disorder detection. Firstly, a representation learning module is introduced with a node information assembly (NIA) mechanism to better extract features from functional connectivity (FC). This mechanism aggregates edge information from both horizontal and vertical directions, effectively assembling node information. Secondly, to generalize the feature across sites, we proposed a site-level feature extraction module that can learn from individual FC data, which circumvents additional prior information. Lastly, an adversarial learning network is proposed as a means of balancing the trade-off between individual classification and site regression tasks, with the introduction of a novel loss function. The proposed method was evaluated on two multi-site fMRI datasets, i.e., Autism Brain Imaging Data Exchange (ABIDE) and ADHD-200. The results indicate that the proposed method achieves a better performance than other related algorithms with the accuracy of 75.56 and 68.92 in ABIDE and ADHD-200 datasets, respectively. Furthermore, the result of the site regression indicates that the proposed method reduces site variability from a data-driven perspective. The most discriminative brain regions revealed by NIA are consistent with statistical findings, uncovering the "black box" of deep learning to a certain extent.

Figures (8)

• Figure 1: Overview of MSalNET. A) Data processing. For each site, FC is calculated via Pearson correlation and site level non-image information such as TR, age range and voxel size are collected for feature selection in site feature learning module. B) Representation learning module. For each FC, we design a representation learning pipeline to grab the individual level features, in the pipeline, NIA aggregates edges information to node through two 1D-CNN kernels sequentially from horizontal and vertical directions. C) Site feature learning module. Based on FC, we construct autoencoder to extract hidden representations of individuals and site features are described by average pooling forming this representation from different site ranges. In site feature learning module, similarity filter step is optional because it may not be executed due to the lack of non-image information. D). Adversarial Learning. A GAN-like adversarial learning network is designed to balance the tradeoff between site regression and disease classification. These two tasks are trained alternately, we designed a new loss function to reduce the site regression and enhance classification at the same time.
• Figure 2: Structure diagram of convolutional neural network. The first convolution layer adopts a horizontal kernel and the result of each convolution can be seen as feature of a ROI. The second convolution layer adopts a vertical kernel to extract whole brain features from each ROI.
• Figure 3: Site feature learning. By utilizing the characteristics of unsupervised training and nonlinear dimensionality reduction of autoencoders, the subject level feature vectors are extracted firstly. Due to the symmetry of the functional connection matrix, the lower trigonometric values, including the main diagonal, are removed. The upper trigonometric part is flattened into a one-dimensional vector and input into AE to obtain the output vector of the encoder, which is the subject level feature vector. To further obtain feature vectors that can reflect the difference information between sites, the average pooling method is used to calculate the mean vector of all subject level feature vectors in each site, which is used as the site feature vector.
• Figure 4: Classification accuracy results of site features in different AE hidden layer dimensions. W/O AE means that the site features are extracted without AE and directly obtained through FC and site average pooling, with a dimension of 19900. The numbers in the figure represent the dimensions of the hidden layer in AE and the dimension of the site feature. A represents ASD, B represents ADHD.
• Figure 5: Classification visualization of brain region contribution with adversarial architecture. We normalize the contribution value of each brain region and select brain regions with a contribution in the range of 0.5-1 for display. The darker the color, the higher the relative importance of the brain region. A for ASD, B for ADHD.