Conceptual

Deep Image Prior with Structured Sparsity for Unsupervised Dynamic MRI Reconstruction

An unsupervised reconstruction method for dynamic (time-resolved) MRI that recovers an image series from undersampled k-space without fully sampled training data. It extends the deep image prior by placing a group-sparsity penalty on frame-specific latent code vectors, forcing an untrained network to discover a low-dimensional manifold that represents temporal variation across frames, showing how an untrained network plus a structured-sparsity prior can replace learned priors and hand-tuned compressed-sensing regularizers.