Longitudinal Image Synthesis for Early Brain Development with Age Awareness

The human brain undergoes dramatic structural and functional changes during the first two postnatal years. Longitudinal magnetic resonance imaging (MRI) enables the observation of dynamic transformations in infant brains over time. However, the longitudinal sampling design and missing time points due to subject dropouts lead to the dispersion of discrete time points. The ability to accurately predict MRI scans of an infant's brain at various time points would aid in dataset construction, offering research potential for applications such as atlas construction and longitudinal trajectory analysis. This task is challenging due to the complex changes in structure and contrast during early brain development. In this work, we propose a three-stage framework with age awareness to predict infant brain MRIs. Our approach focuses on modeling three key aspects of early brain development at each stage: volumetric expansion, tissue growth, and myelination. We validated our technique on longitudinal MRI data from the Baby Connectome Project, consisting of 491 T1-weighted images. Quantitative analysis demonstrates that our model synthesizes realistic infant brain MRIs at target ages with high anatomical accuracy. Our predicted images exhibit appropriate structural alignment and tissue contrast reflective of the age. The code will be available upon acceptance.