A Deep Learning Approach to Longitudinal Infant MRI Synthesis

Motivation: Longitudinal MRI studies of infants often face missing time points due to dropout, motion, and other issues. We propose a method to synthesize missing infant MRIs and aid early brain development research. Goal(s): We present a deep learning framework that synthesizes T1w images for any target age (0-2 yrs) from a different source age. Approach: Our three-stage framework includes global affine transformation, local deformation, and contrast synthesis, guided by temporal attention for realistic age-specific MRIs. Results: Our method outperforms others in image quality and segmentation tasks, and aids in constructing developmental trajectories. Impact: This framework enables accurate tracking of infant brain development by filling missing MRI data, aiding in the creation of developmental atlases, and supporting early detection of disorders. It may thus advance both neurodevelopmental research and clinical interventions.