Modeling the acquisition shift between axial and sagittal MRI for diffusion superresolution to enable axial spine segmentation

Spine MRIs are usually acquired in highly anisotropic 2D axial or sagittal slices. Vertebra structures are not fully resolved in these images, and multi-image superresolution by aligning scans to pair them is difficult due to partial volume effects and inter-vertebral movement during acquisition. Hence, we propose an unpaired inpainting superresolution algorithm that extrapolates the missing spine structures. We generate synthetic training pairs by multiple degradation functions that model the data shift and acquisition errors between sagittal slices and sagittal views of axial images. Our method employs modeling of the k-space point spread function and the interslice gap. Further, we imitate different MR acquisition challenges like histogram shifts, bias fields, interlace movement artifacts, Gaussian noise, and blur. This enables the training of diffusion-based superresolution models on scaling factors larger than 6$\times$ without real paired data. The low z-resolution in axial images prevents existing approaches from separating individual vertebrae instances. By applying this superresolution model to the z-dimension, we can generate images that allow a pre-trained segmentation model to distinguish between vertebrae and enable automatic segmentation and processing of axial images. We experimentally benchmark our method and show that diffusion-based superresolution outperforms state-of-the-art super-resolution models.