MAD-Net: Morphometric-Attentive Diffusion Network for Predicting Longitudinal Infant Brain Functional Connectivity

Resting-state functional MRI (rs-fMRI) data analysis provides essential insights into early neurodevelopment through longitudinal assessment of functional connectivity (FC) patterns in infant brains, which may help uncover critical biomarkers for developmental monitoring. However, due to challenges in acquiring high-quality functional MRI (fMRI) data in infants, such as strong motion artifacts, short scan durations, and participant compliance, longitudinal FC of infants remain scarce, which significantly hampers the capacity to systematically investigate early functional brain development. To address this challenge, we propose MAD-Net, a novel diffusion model that predicts longitudinal FC from morphometric features derived from structural MRI (sMRI). Our framework integrates classifier-free guidance with a cross-modal attention mechanism, enabling the dynamic fusion of morphometric features and developmental age constraints during the diffusion process. A shared triplet encoder learns robust feature representations from longitudinal data, while a U-Net-based architecture ensures precise conditioning on individual morphometry and target age. We evaluate MAD-Net on 386 longitudinal infant fMRI scans and demonstrate its superior performance in FC prediction compared to state-of-the-art methods. By integrating diffusion-based learning, structural priors, and age-dependent constraints, MAD-Net represents a significant advancement in neuroimaging-based functional network reconstruction. The code is available at https://github.com/IPMI-NWU/MAD-Net.