Volume reconstruction and view synthesis from 2D projection MRI using 3D Gaussian Splatting

Motivation: Volumetric MRI often cannot be directly achieved because of contrast requirements, subject motion, etc. SPI provides a motion-robust 3D acquisition, and acceleration can be achieved with novel methods. Goal(s): To reconstruct volumetric MR images from a sparse set of 2D projections. Approach: We demonstrate an alternative novel method for volume synthesis using multiple spiral imaging projections based on 3D Gaussian Splatting, a recently developed method in computer vision for image and surface rendering from a sparse set of camera inputs. Results: In the simulated noiseless case, this method is accurate and performs well with up to 8-fold undersampling, compared to fully-sampled gridding reconstruction. Impact: We demonstrate a new approach for MRI volume synthesis based on multiple projections (e.g. spiral projection imaging), using noise-less neuroimaging simulations. This has the potential to enable shorter scans and/or better imaging in low signal-to-noise settings.