ADHD Classification Model Based on Functional Brain Network and Graph Feature Learning
-
摘要: 功能脑网络(FBN)在精神障碍诊断中广泛应用, 但传统构建方法缺乏与下游任务的互动性, 限制了模型性能; 且图神经网络多层堆叠易导致节点特征过度平滑难以提取深层特征. 为此, 提出端到端的自适应聚合功能网络模型, 通过大脑感兴趣区域(ROI)感知汇聚层, 利用自注意力机制动态构建FBN并学习节点特征, 增强了模型与子任务的交互能力. 同时引入节点池化机制筛选显著ROI, 进而推断出对于子任务较为重要的ROI. 该方法应用于注意力缺陷多动障碍(ADHD)的分类实验中, 实验结果表明该方法提高了ADHD的分类准确率, 对实验结果的解释性分析也验证了该方法的有效性.Abstract: Functional brain network (FBN) is widely used in the diagnosis of psychiatric disorders. However, traditional construction methods lack interactivity with downstream tasks, limiting model performance; Additionally, multi-layer stacking of graph neural networks may lead to over-smoothing of node features, making it difficult to extract deep-level features. To address these issues, we propose an end-to-end adaptive aggregated functional network model. This model dynamically constructs FBN and learns node features through a brain region-of-interest (ROI) awareness aggregation layer, which integrates a self-attention mechanism. This method enhances the model＇s interaction capability with sub-tasks. Simultaneously, a node pooling mechanism is introduced to filter salient ROI, thereby inferring ROI that are critical for specific sub-tasks. The proposed method is applied to classification experiments of attention deficit hyperactivity disorder (ADHD). Experimental results demonstrate the method improves the ADHD classification accuracy, and the interpretability analysis of experimental results further validates the effectiveness of the proposed method.1)
1 1http://fcon_1000.projects.nitrc.org/indi/adhd200/ 2https://neurobureau.projects.nitrc.org/ADHD200/Introduction.html/ 2)2 2https://neurobureau.projects.nitrc.org/ADHD200/Introduction.html/ -
图 2 ROI感知汇聚池化层结构示意图
Fig. 2 Schematic diagram of the ROI awareness aggregated pooling layer structure
图 4 NYU数据集中ADHD类别和HC类别构建FBN的可视化
Fig. 4 Visualization of FBN constructed for ADHD and HC categories in the NYU dataset
图 5 NYU数据集中与ADHD分类较为密切的大脑区域
Fig. 5 Brain regions closely related to ADHD classification in the NYU dataset
图 6 在NYU数据集上不同$r_k$取值对模型性能的影响
Fig. 6 The impact of different $r_k$ values on model performance in the NYU dataset.
图 7 在NYU数据集上不同$M$取值对模型性能的影响
Fig. 7 The impact of different $M$ values on model performance in the NYU dataset.
表 1 论文中使用的符号及其含义
Table 1 Symbols and their meanings used in the paper
符号 含义 $ D^{(l)} $ 第$ l $层特征向量的维数 $ M $ AAFNBlock堆叠的层数 $ N $ 第$ l $层输入的局部特征个数 $ n $ 样本个数 $ E $ 对角元素全为1的单位矩阵 $ P $ 基于正弦−余弦函数生成的位置编码矩阵 $ ROI_{i} $ 编号为$ i $的大脑感兴趣区域 $ CLS^{(l)} $ 第$ l $层输出的全局特征嵌入矩阵 $ cls^{(l)} $ 第$ l $层输出的全局特征嵌入向量 $ H^{(l)} $ 执行第$ l $层ROI池化后的局部特征嵌入矩阵 $ r_{k} $ 每层ROI池化层TopK参数$ k $衰减率 $ \tilde H^{(l)} $ 执行第$ l $层ROI池化前的局部特征嵌入矩阵 $ h_i^{(l)} $ 执行第$ l $层汇聚后与$ ROI_{i} $相关的局部特征向量, $ h_i^{(l)} \in {\bf{R}}^{D^{(l)}} $ $ \tilde h_i^{(l)} $ 执行第$ l $层汇聚前与$ ROI_{i} $相关的局部特征向量, $ \tilde h_i^{(l)} \in {\bf{R}}^{D^{(l)}} $ $ h_{j,\;i} $ $ h_i^{(0)} $向量掩码前第$ j $维的值, $ h_i^{(0)} \in H^{(0)} $ $ p_{j,\;i} $ 解码后与$ h_{j,\;i} $对应的值 $ \text{Me-pooling} $ 均值池化操作 $ RelationMap^{(l)} $ 第$ l $层局部特征之间以及与全局特征之间的相关性矩阵 $ \alpha_G $ 全局特征与局部特征之间的相关性向量 $ k $ 经过ROI池化层后保留的节点个数 $ \lambda $ 损失函数相关超参数 $ W $ 可学习的卷积核 $ w $ 计算嵌入向量相关性的可学习参数 $ {\rm{w}} $ 模型中所有可学习的参数 
下载: 导出CSV
表 2 数据分布
Table 2 Data distribution
数据集 训练数据集 测试数据集 ADHD HC ADHD HC KKI 41 104 9 28 NYU 256 190 62 60 NI 58 51 14 16 PKU 164 228 40 58
下载: 导出CSV
表 3 模型超参数
Table 3 Model Hyperparameters
超参数 KKI NYU NI PKU 初始学习率 $2.25 \times 10^{-2}$ $2.25 \times 10^{-2}$ $2.25 \times 10^{-2}$ $2.2 \times 10^{-2}$ 训练轮次 100 100 100 100 每轮批量数 15 32 32 32 $ \lambda_{1} $ $5 \times 10^{-4}$ $5 \times 10^{-4}$ $5 \times 10^{-4}$ $5 \times 10^{-4}$ $ \lambda_{2} $ $1.875 \times 10^{-3}$ $1.875 \times 10^{-3}$ $1.875 \times 10^{-3}$ $1.875 \times 10^{-3}$ $ r_k $ 0.5 0.5 0.5 0.5 $ M $ 2 2 2 2
下载: 导出CSV
表 4 模型性能(%)
Table 4 Model Performance (%)
数据集 KKI NYU PKU NI ACC 92.7 86.8 86.0 92.3 SP 96.3 82.0 90.5 94.4 SE 79.9 90.4 80.4 90.3 PR 89.8 87.1 86.3 94.8 F1 83.5 88.7 83.0 92.2
下载: 导出CSV
表 5 ADHD-200数据集上最新方法与所提出模型的分类准确率比较(%)
Table 5 Comparison of classification accuracy on ADHD-200 dataset (%)
下载: 导出CSV
表 6 全局特征提取方法对模型性能准确率、特异度和敏感度的影响(%)
Table 6 Impact of global feature extraction methods on model accuracy, specificity and sensitivity (%)
数据集 ACC SP SE KKI NYU PKU NI KKI NYU PKU NI KKI NYU PKU NI $ {\rm{Max}} $ 读出层 87.5 83.7 78.6 85.7 94.4 76.9 83.9 90.9 61.5 88.9 71.2 81.6 $ {\rm{Mean}} $ 读出层 86.4 84.0 79.2 87.0 93.1 80.0 87.9 95.3 63.3 87.0 67.4 79.4 本文方法 92.7 86.8 86.0 92.3 96.3 82.0 90.5 94.4 79.9 90.4 80.4 90.3
下载: 导出CSV
表 7 全局特征提取方法对模型性能精确度、F1的影响(%)
Table 7 Impact of global feature extraction methods on model precision and F1 (%)
数据集 PR F1 KKI NYU PKU NI KKI NYU PKU NI $ {\rm{Max}} $ 读出层 83.9 83.8 77.4 90.6 67.9 86.2 74.0 84.7 $ {\rm{Mean}} $ 读出层 72.9 85.1 80.3 95.0 64.2 86.0 73.1 85.9 本文方法 89.8 87.1 86.3 94.8 83.5 88.7 83.0 92.2
下载: 导出CSV
-
[1] Wang X H, Jiao Y, Li L. Identifying individuals with attention deficit hyperactivity disorder based on temporal variability of dynamic functional connectivity. Scientific Report, 2018, 8(1): Article No. 11789 doi: 10.1038/s41598-018-30308-w [2] Wender P H, Wolf L E, Wasserstein J. Adults with ADHD: An overview. Annals of the New York Academy of Sciences, 2001, 931(1): 1−16 doi: 10.1111/j.1749-6632.2001.tb05770.x [3] Cicek G, Akan A. Deep learning approach versus traditional machine learning for ADHD classification. In: Proceedings of the 2021 Medical Technologies Congress (TIPTEKNO). Kundu, Antalya, Turkey: IEEE, 2021. 1–4 [4] Mao Z, Su Y, Xu G, Wang X, Huang Y, Yue W, et al. Spatio-temporal deep learning method for ADHD fMRI classification. Information Sciences, 2019, 499: 1−11 doi: 10.1016/j.ins.2019.05.043 [5] Riaz A, Asad M, Alonso E, Slabaugh G. DeepFMRI: End-to-end deep learning for functional connectivity and classification of ADHD using fMRI. Journal of Neuroscience Methods, 2020, 335: Article No. 108506 doi: 10.1016/j.jneumeth.2019.108506 [6] van den Heuvel M P, Pol H E H. Exploring the brain network: A review on resting-state fMRI functional connectivity. European Neuropsychopharmacology, 2010, 20(8): 519−534 doi: 10.1016/j.euroneuro.2010.03.008 [7] Jia H, Wu G, Wang Q, Shen D. ABSORB: Atlas building by self-organized registration and bundling. NeuroImage, 2010, 51(3): 1057−1070 doi: 10.1016/j.neuroimage.2010.03.010 [8] Jia H, Yap P T, Shen D. Iterative multi-atlas-based multi-image segmentation with tree-based registration. NeuroImage, 2012, 59(1): 422−430 doi: 10.1016/j.neuroimage.2011.07.036 [9] Riaz A, Asad M, Alonso E, Slabaugh G. Fusion of fMRI and non-imaging data for ADHD classification. Computerized Medical Imaging and Graphics, 2018, 65: 115−128 doi: 10.1016/j.compmedimag.2017.10.002 [10] Dai D, Wang J, Hua J, He H. Classification of ADHD children through multimodal magnetic resonance imaging. Frontiers in Systems Neuroscience, 2012, 6: Article No. 63 [11] Liu M, Zhang J, Adeli E, Shen D. Joint classification and regression via deep multi-task multi-channel learning for Alzheimer’s disease diagnosis. IEEE Transactions on Biomedical Engineering, 2018, 66(5): 1195−1206 [12] Zhao K, Duka B, Xie H, Oathes D J, Calhoun V, Zhang Y. A dynamic graph convolutional neural network framework reveals new insights into connectome dysfunctions in ADHD. NeuroImage, 2022, 246: Article No. 118774 doi: 10.1016/j.neuroimage.2021.118774 [13] Zhang S, Chen X, Shen X, Ren B, Yu Z, Yang H, et al. A-GCL: Adversarial graph contrastive learning for fMRI analysis to diagnose neurodevelopmental disorders. Medical Image Analysis, 2023, 90: Article No. 102932 doi: 10.1016/j.media.2023.102932 [14] Bellec P, Chu C, Chouinard-Decorte F, Benhajali Y, Margulies D S, Craddock R C. The neuro bureau ADHD-200 preprocessed repository. NeuroImage, 2017, 144: 275−286 doi: 10.1016/j.neuroimage.2016.06.034 [15] Zuo Q, Wu H, Chen C P, Lei B, Wang S. Prior-guided adversarial learning with hypergraph for predicting abnormal connections in Alzheimer's disease. IEEE Transactions on Cybernetics, 2024, 54(6): 3652−3665 doi: 10.1109/TCYB.2023.3344641 [16] Zhang J, Zhou L, Wang L, Liu M, Shen D. Diffusion kernel attention network for brain disorder classification. IEEE Transactions on Medical Imaging, 2022, 41(10): 2814−2827 doi: 10.1109/TMI.2022.3170701 [17] Kreuzer D, Beaini D, Hamilton W, Létourneau V, Tossou P. Rethinking graph transformers with spectral attention. Advances in Neural Information Processing Systems, 2021, 34: 21618−21629 [18] Tzourio-Mazoyer N, Landeau B, Papathanassiou D, Crivello F, Etard O, Delcroix N, et al. Automated anatomical labeling of activations in SPM using a macroscopic anatomical parcellation of the MNI MRI single-subject brain. NeuroImage, 2002, 15(1): 273−289 doi: 10.1006/nimg.2001.0978 [19] Saad J F, Griffiths K R, Korgaonkar M S. A systematic review of imaging studies in the combined and inattentive subtypes of attention deficit hyperactivity disorder. Frontiers in Integrative Neuroscience, 2020, 14: Article No. 31 doi: 10.3389/fnint.2020.00031 [20] Cangea C, Veličković P, Jovanović N, Kipf T, Liò P. Towards sparse hierarchical graph classifiers. arXiv: 1811.01287, 2018. [21] Yan C, Zang Y. DPARSF: A MATLAB toolbox for “pipeline” data analysis of resting-state fMRI. Frontiers in systems neuroscience, 2010, 4: Article No. 1377 [22] Saad J F, Griffiths K R, Kohn M R, Clarke S, Williams L M, Korgaonkar M S. Regional brain network organization distinguishes the combined and inattentive subtypes of attention deficit hyperactivity disorder. NeuroImage: Clinical, 2017, 15: 383–390 [23] Suskauer S J, Simmonds D J, Caffo B S, Denckla M B, Pekar J J, Mostofsky S H. fMRI of intrasubject variability in ADHD: Anomalous premotor activity with prefrontal compensation. Journal of the American Academy of Child & Adolescent Psychiatry, 2008, 47(10): 1141−1150 [24] Solanto M. V, Schulz K P, Fan J, Tang C Y, Newcorn J H. Event-related FMRI of inhibitory control in the predominantly inattentive and combined subtypes of ADHD. Journal of Neuroimaging, 2009, 19(3): 205−212 doi: 10.1111/j.1552-6569.2008.00289.x [25] Dey S, Rao A R, Shah M. Exploiting the brain's network structure in identifying ADHD subjects. Frontiers in systems neuroscience, 2012, 6: Article No. 75 [26] Zou L, Zheng J, Miao C, Mckeown M J, Wang Z J. 3D CNN based automatic diagnosis of attention deficit hyperactivity disorder using functional and structural MRI. IEEE Access, 2017, 5: 23626−23636 doi: 10.1109/ACCESS.2017.2762703 -
计量
- 文章访问数: 23
- HTML全文浏览量: 24
- 被引次数: 0
