Lesion Segmentation of Neonatal Diffusion MRI Under Simulated K-space Undersampling

Accurate segmentation of Hypoxic Ischemic Encephalopathy (HIE) lesions in neonatal diffusion MRI remains challenging due to the small, diffuse nature of the lesions and the constraints of accelerated imaging. Detecting such tiny lesions becomes particularly difficult when MRIs are acquired at very low spatial resolutions. In this study, we address the task of lesion segmentation from low-resolution MR images by simulating realistic clinical conditions through 4-fold equispaced undersampling of Apparent Diffusion Coefficient (ADC) and Z-score ADC (Z_ADC) volumes from the BONBID-HIE dataset. Paired aliased images are generated via inverse Fourier reconstruction, and three segmentation models are trained under these settings: (i) a baseline SwinUNETR, (ii) a DCT-enhanced model incorporating global frequency representations, and (iii) a variant integrating localized (block-based) DCT into early encoder layers. Experimental results demonstrate that lesion segmentation from undersampled images is feasible, and that incorporating frequency-domain priors significantly improves performance. Our findings highlight the advantages of combining transformer-based architectures with frequency-aware augmentation for robust neonatal brain lesion segmentation under limited k-space sampling.