Abstract
Deep learning methods for accelerated MRI achieve state-of-the-art resultsbut largely ignore additional speedups possible with noncartesian samplingtrajectories. To address this gap, we created a generative diffusionmodel-based reconstruction algorithm for multi-coil highly undersampled spiralMRI. This model uses conditioning during training as well as frequency-basedguidance to ensure consistency between images and measurements. Evaluated onretrospective data, we show high quality (structural similarity > 0.87) inreconstructed images with ultrafast scan times (0.02 seconds for a 2D image).We use this algorithm to identify a set of optimal variable-density spiraltrajectories and show large improvements in image quality compared toconventional reconstruction using the non-uniform fast Fourier transform. Bycombining efficient spiral sampling trajectories, multicoil imaging, and deeplearning reconstruction, these methods could enable the extremely highacceleration factors needed for real-time 3D imaging.