TADM-CGAN: a resting-state to task activation map prediction framework using temporal attention-driven diffusion models and conditional generative adversarial networks

Abstract

Predicting task-induced brain activation from resting-state fMRI (rs-fMRI) remains a significant challenge in computational neuroimaging, primarily due to the difficulty in simultaneously modeling the detailed temporal evolution and high spatial resolution of intrinsic neural activity. Most existing literature relies on parcel-based modeling using spatial functional connectivity features, neglecting the long-range temporal interactions and nonlinear fluctuations in rs-fMRI signals. To overcome these limitations, we introduce TADM-CGAN, a two-stage, grayordinate-level cascaded architecture that infers task activation maps directly from rs-fMRI time series, fully utilizing both temporal and spatial characteristics. In the first stage, a multi-head temporal attention-driven diffusion model (TADM) is employed to generate compact temporal embeddings for each grayordinate, capturing dependencies across the entire rs-fMRI time series. These unique temporal features then serve as inputs to 59,412 time series regression models, enabling highly localized, grayordinate-specific prediction of preliminary activation values. In the second stage, a principal component analysis (PCA)-conditioned conditional generative adversarial network (PCA-CGAN) is introduced, where PCA constrains adversarial refinement to a low-rank, biologically meaningful subspace, while the generator with the proposed activation fidelity loss (AFL) reduces noise and sharpens spatial details for the improved predictions. This proposed cascaded framework consistently outperforms prior task activation map prediction methods across a diverse set of task contrasts and datasets, underscoring its robustness and generalizability in practical applications.