Accelerating Longitudinal MRI using Prior Informed Latent Posterior Sampling (PIPS)

Motivation: Long scan times persist longitudinal MRI studies, yet lack good reconstruction priors for accelerated parallel imaging. Goal(s): Substantially and scaleably reduce scantime by leveraging prior scans of the same patient in a data-consistent deep learning reconstruction. Approach: Propose Prior Informed Posterior Sampling (PIPS) with latent diffusion. We train such a model on unlabeled image data, eliminating the need for a dataset of either k-space measurements or paired longitudinal scans required of other learning-based methods. Results: Outperform baselines in effectively utilizing prior information without over-biasing prior consistency. Validated on an open-source dataset of healthy patients, as well as several longitudinal cases with clinical differentials. Impact: We propose an unsupervised prior conditioning method to further accelerate MRI for longitudinal studies. Our method is both scalable and generalizable, as it does not require sequential k-space for training and enforces data consistency throughout the reconstruction.