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

Area of Science:

• Neuroscience
• Medical Imaging
• Artificial Intelligence

Background:

• Resting-state functional MRI (rs-fMRI) is vital for assessing infant brain development and identifying biomarkers.
• Acquiring high-quality infant fMRI data is challenging due to motion, short scans, and compliance issues, limiting longitudinal studies.
• Existing methods struggle to reconstruct longitudinal functional connectivity (FC) in infants.

Purpose of the Study:

• To develop a novel method for predicting longitudinal functional connectivity (FC) in infant brains.
• To address the scarcity of longitudinal infant fMRI data by leveraging structural MRI (sMRI) information.
• To improve early neurodevelopmental monitoring through advanced neuroimaging analysis.

Main Methods:

• Proposed MAD-Net, a diffusion model integrating classifier-free guidance and cross-modal attention.
• Utilized morphometric features from sMRI to predict longitudinal FC.
• Employed a shared triplet encoder for robust longitudinal feature representation and a U-Net architecture for precise conditioning.

Main Results:

• MAD-Net demonstrated superior performance in predicting longitudinal FC compared to state-of-the-art methods.
• The model effectively fused morphometric features and developmental age constraints.
• Evaluated on 386 longitudinal infant fMRI scans, showing significant advancements in functional network reconstruction.

Conclusions:

• MAD-Net offers a significant advancement in reconstructing infant brain functional networks using structural priors.
• The model overcomes limitations of scarce longitudinal infant fMRI data.
• This approach holds promise for enhancing early neurodevelopmental monitoring and biomarker discovery.