The temporal generalized autocalibrating partially parallel acquisitions (TGRAPPA) algorithm for parallel MRI was modified for real-time low latency imaging in interventional procedures using image domain, B(1)-weighted reconstruction. GRAPPA coefficients were calculated in k-space, but applied in the image domain after appropriate transformation. Convolution-like operations in k-space were thus avoided, resulting in improved reconstruction speed. Image domain GRAPPA weights were combined into composite unmixing coefficients using adaptive B(1)-map estimates and optimal noise weighting. Images were reconstructed by pixel-by-pixel multiplication in the image domain, rather than time-consuming convolution operations in k-space. Reconstruction and weight-set calculation computations were parallelized and implemented on a general-purpose multicore architecture. The weight calculation was performed asynchronously to the real-time image reconstruction using a dedicated parallel processing thread. The weight-set coefficients were computed in an adaptive manner with updates linked to changes in the imaging scan plane. In this implementation, reconstruction speed is not dependent on acceleration rate or GRAPPA kernel size.
Trial registration: ClinicalTrials.gov NCT00720460.