BDHT: Generative AI Enables Causality Analysis for Mild Cognitive Impairment

TL;DR

The paper tackles the challenge of reliably estimating directed effective connectivity (EC) from fMRI in mild cognitive impairment (MCI), where preprocessing variability can bias results. It introduces BDHT, a diffusion-based generative framework that conditions the reverse denoising process on both rough fMRI-derived samples and structural connectivity from DTI to produce clean ROI time series and EC estimates end-to-end. The architecture combines a hierarchical denoising transformer (with ConAttention alignment and GraphConFormer blocks) within a DDPM, enabling multi-scale topological feature learning and structure-function integration, which yields improved denoising quality and EC accuracy. On the ADNI dataset, BDHT achieves superior reconstruction fidelity and classification performance (NC vs EMCI/LMCI) and identifies altered directional connections and key ROIs that could serve as biomarkers for MCI progression and treatment guidance.

Abstract

Effective connectivity estimation plays a crucial role in understanding the interactions and information flow between different brain regions. However, the functional time series used for estimating effective connectivity is derived from certain software, which may lead to large computing errors because of different parameter settings and degrade the ability to model complex causal relationships between brain regions. In this paper, a brain diffuser with hierarchical transformer (BDHT) is proposed to estimate effective connectivity for mild cognitive impairment (MCI) analysis. To our best knowledge, the proposed brain diffuser is the first generative model to apply diffusion models to the application of generating and analyzing multimodal brain networks. Specifically, the BDHT leverages structural connectivity to guide the reverse processes in an efficient way. It makes the denoising process more reliable and guarantees effective connectivity estimation accuracy. To improve denoising quality, the hierarchical denoising transformer is designed to learn multi-scale features in topological space. By stacking the multi-head attention and graph convolutional network, the graph convolutional transformer (GraphConformer) module is devised to enhance structure-function complementarity and improve the ability in noise estimation. Experimental evaluations of the denoising diffusion model demonstrate its effectiveness in estimating effective connectivity. The proposed model achieves superior performance in terms of accuracy and robustness compared to existing approaches. Moreover, the proposed model can identify altered directional connections and provide a comprehensive understanding of parthenogenesis for MCI treatment.

Figures (14)

• Figure 1: The schematic diagram of the proposed BDHT model, including the diffusive process (dashed arrows) and the denoising process (solid arrows). The two conditions are the fMRI and DTI.
• Figure 2: The detailed network structure of one denoising step in the proposed BDHT model. In the diffusive process, the empirical sample $\boldsymbol{H}_0$ is gradually transformed into the Gaussian sample $\boldsymbol{H}_T$ by adding noise at each successive step. In the denoising process, the rough sample $\boldsymbol{F}$ and structural connectivity $\mathcal{G}$ are input as conditions to guide the hierarchical denoising transformer to generate the clean sample and effective connectivity.
• Figure 3: The structure of the ConAttention Alignment. The inputs are the rough sample $\boldsymbol{F}$ and intermediate noisy sample $\boldsymbol{H}_t$, and the output is the fused feature.
• Figure 4: The computing structure of the GraphConFormer block. The input is the fused feature, and the output updates the input and shares the same dimension. The $\mathcal{G}$ guides the topological feature learning in the GCN layers.
• Figure 5: An example of denoising the conditional sample at different steps for the first ROI in the AAL90 atlas. The top arrow is the conditional time series calculated by the non-parametric dot product, and the bottom row is the empirical time series computed by the GRETNA software.