Speaker
Description
Recovering high-fidelity 3D volumes from sparse or degraded 2D images is a fundamental challenge in magnetic resonance imaging with broad clinical applications. Implicit neural representations (INRs) offer a resolution-agnostic solution to modelling a volume directly from a scanner’s coordinate system. This allows for continuous 3D representations to be built from arbitrarily oriented sets of imaging slices, as the exact position and orientation of every voxel can be derived from DICOM headers. Modelling volumes directly on the spatial geometry allows us to impose physical constraints into the learning process. We can employ the point-spread function to accurately model the acquisition physics of a given anisotropic voxel. Moreover, inter-slice subject motion can be corrected in an unsupervised manner by expressing a slice’s rigid transformation as optimizable translation and rotation parameters. These methods are also capable of handling multi-subject cohorts, allowing for statistical correlations on tissue position to be exploited across a population. In our recent works, we apply these methods to a multi-subject cohort of 2D+time cardiac CINE datasets. We demonstrate that, by training on intensity-segmentation pairs, these representations can interpolate 3D+time volumes at any desired resolution. At test-time deriving a new subject’s representation from imaging information alone provides free complimentary segmentation labels. Ongoing work includes applying further physical constraints. The time dimension can be framed under a velocity-based deformation process, enforcing physically-grounded tissue motion. Furthermore, exchanging the neural representation in favor of Gaussian primitives allow us to formulate analytically-computable versions of the aforementioned constraints, drastically reducing compute costs.