Implicit neural representation for dynamic imaging

Dynamic imaging plays a fundamental role in studying various biological phenomena but faces two challenges: data incompleteness and computational burden. This contribution investigates the use of implicit neural representations (INRs) to address both. Using INRs, a dynamic image reconstruction method is proposed, where a continuous mapping from spatiotemporal coordinates to a scalar value representing the object is learned directly from measurement data. As such, it is fundamentally different from other learning methods requiring large datasets of training images. The feasibility of the proposed framework is illustrated with an application to dynamic image reconstruction from undersampled circular Radon transform data.