












摘" 要: 針對動態因果模型(DCM)在分析大腦有效連接時面臨的高計算成本問題,提出一種結合廣義線性模型(GLM)和稀疏DCM的算法,即廣義稀疏DCM(GSD)算法。該算法在以下三個方面進行優化:首先,利用傅里葉變換的對稱性將頻域DCM的復數計算轉換為實數計算,降低計算復雜度;其次,應用GLM和濾波技術減少觀測信號的干擾,提高參數估計的準確度;最后,定義新的代價函數來優化變分推斷參數和濾波器參數,進一步提升參數估計的精度。該研究采用兩組公開的功能核磁共振成像(fMRI)數據對GSD算法進行驗證,包括仿真的史密斯小世界網絡數據和運動與注意力實測數據。實驗結果表明,GSD算法在保持與傳統方法相近的參數估計性能的同時,能將計算時間降低50%以上。該研究成果為平衡模型的解釋力和計算效率提供了新的視角,有望推動DCM在更廣泛領域的應用。
關鍵詞: 廣義線性模型; 動態因果模型; 稀疏算法; 有效連接; 計算復雜度; 模型解釋力
中圖分類號: TN919?34; TP301.6" " " " " " " " " "文獻標識碼: A" " " " " " " " " " 文章編號: 1004?373X(2025)07?0146?09
Fast dynamic causal modeling regression method for fMRI
HU Xinhang1, WU Haifeng1, 2, 3, ZENG Yu1, 2, 3
(1. School of Electrical and Information Engineering, Yunnan Minzu University, Kunming 650504, China;
2. Yunnan Provincial Key Laboratory of Unmanned Autonomous Systems, Kunming 650504, China;
3. Yunnan Provincial Colleges and Universities Intelligent Sensor Network and Information System Technology Innovation Team,
Yunnan Minzu University, Kunming 650504, China)
Abstract: In view of the high computational cost faced by dynamic causal modeling (DCM) when analyzing brain effective connectivity, an algorithm that combines the generalized linear model (GLM) and sparse DCM, namely the combined generalize?linear?model and sparse DCM (GSD) algorithm, is proposed. The algorithm is optimized in the following three aspects: firstly, the symmetry of Fourier transform is used to convert the complex number calculation of frequency domain DCM into real number calculation, reducing the computational complexity; secondly, GLM and filtering techniques are applied to reduce the interference of observed signals and improve the accuracy of parameter estimation; finally, a new cost function is defined to optimize the variational inference parameters and filter parameters to further improve the accuracy of parameter estimation. In this study, two sets of public functional magnetic resonance imaging (fMRI) data are used to verify the GSD algorithm, including the simulated Smith small?world network data and the motion and attention measured data. The experimental results show that the GSD algorithm can reduce the calculation time by more than 50% while maintaining parameter estimation performance similar to that of the traditional methods. The research results of this paper provide a new perspective for balancing the explanatory power and computational efficiency of the model, and are expected to promote the application of DCM in a wider range of fields.
Keywords: GLM; DCM; sparse algorithm; effective connectivity; computational complexity; model explanatory power
0" 引" 言
隨著腦成像技術的不斷發展,出現了許多結合功能核磁共振成像(Functional Magnetic Resonance Imaging, fMRI)技術來研究有效連接的方法,例如結構方程模型、多變量回歸模型、格蘭杰因果模型(Granger Causality Modeling, GCM)和動態因果模型(Dynamic Causal Modeling, DCM)等[1],其中GCM和DCM最具代表性。DCM基于動態系統理論和狀態空間模型[2],不僅能夠更全面地考慮大腦網絡的動態調節過程,也可以根據神經科學先驗知識對腦活動進行解釋,這使得DCM成為了研究有效連接更主流的方法,并涌現出了許多與之相結合的新方法[3?5]。……