Accelerating MRI With Longitudinally-Informed Latent Posterior Sampling.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Reconstruction

Background:

• Longitudinal MRI is crucial but leveraging prior scans for reconstruction is difficult.
• Existing deep learning models require paired longitudinal data, which is scarce.
• Substantial anatomical changes between scans pose challenges for traditional methods.

Purpose of the Study:

• To accelerate MRI acquisition by incorporating prior scans into the reconstruction process.
• To develop a reconstruction framework that does not require longitudinally paired training data.
• To introduce a novel open-access clinical dataset for longitudinal MRI research.

Main Methods:

• A diffusion-model-based reconstruction framework was proposed.
• The model was trained using standalone images, treating all timepoints as samples from the same distribution.
• Prior scans (DICOM format) were integrated at inference to guide follow-up scan reconstruction.

Main Results:

• The proposed method outperformed longitudinal and non-longitudinal baselines in accelerated Cartesian imaging.
• Image quality improved by up to 10% in SSIM and 2 dB in PSNR in regions similar to prior scans.
• The method showed robustness to anatomical changes and misregistration compared to longitudinal baselines.

Conclusions:

• Prior scans can be effectively integrated with diffusion-based reconstruction for improved MRI.
• The approach enhances image quality and enables greater scan acceleration.
• This method circumvents the need for extensive longitudinally paired training datasets.