Generation of Anatomy-Realistic 4d Infant Brain Atlases with Tissue Maps Using Generative Adversarial Networks

Brain development during infancy is complex and dynamic, making it crucial yet challenging to generate longitudinal 4D atlases with densely sampled time points that accurately reflect early developmental patterns. Accordingly, learning-based methods have been proposed to build longitudinally continuous atlases. However, these methods do not generate tissue segmentation maps (TSMs) along with atlases, which are critical in atlas-guided segmentation and spatial normalization. In addition, intensity images, which exhibit low and dynamic contrast during infancy, provide little guidance on generating anatomically realistic features compared to TSMs, which provide explicit information on the relative position of different brain tissues. To this end, we propose a continuous 4D atlas learning framework that constructs sharp and anatomically realistic templates equipped with accurate TSMs. Specifically, our work focuses on (1) exploiting structural information in the available TSMs to encourage anatomically more realistic intensity templates, and (2) incorporating an affine network that re-scales the generated atlases in the same atlas space to the corresponding age-specific spaces. The proposed framework is employed to construct high-fidelity 4D infant brain volumetric atlases based on 699 infant brain MRI scans from 322 subjects from birth to 73 months. Experimental results show that our 4D atlases contain anatomically meaningful features with sharper structural details, compared to 4D atlases generated by other methods.