T2T-Bridge: Direct Diffusion Bridge Model for Thick-to-Thin Slice Infant MRI Reconstruction

Accurate diagnosis and prognosis of pediatric brain diseases require high-quality MRI data. However, due to low tolerance and cooperation, pediatric neuroimaging is extremely challenging. In most clinical practice, only 2D thick-slice MRI are acquired. This type of data suffers poor through-plane resolution, losing detailed morphological and geometrical information that is often provided by 3D high-resolution (thin-slice) MRI. AI-based super-resolution of the pediatric clinical MRI holds promise for more thorough and accurate diagnosis, but can be very difficult compared to adult MRI super-resolution due to smaller brain, variable contrast, and low SNR. Current AI models could fail to retain image fidelity on this task. In this paper, we propose T2T-Bridge that learns an optimal transport path between thick-and thin-slice MRI distributions using a diffusion architecture. Specifically, we progressively reconstruct thin-slice images from thick-slice images, rather than starting from random Gaussian noise. This preserves valuable structural information contained in the original 2D MRI. This gradual resolution enhancement is also more viable than single mapping networks. We introduce anatomical priors to strengthen structural consistency and incorporate age information to adapt image style, enabling super-resolution for infant images across all age groups. Our experiments on a clinical infant dataset demonstrate the high-fidelity super-resolution performance of T2T-Bridge.