LDPM: Towards undersampled MRI reconstruction with MR-VAE and Latent Diffusion Prior.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• Diffusion models show promise for image reconstruction, but pixel-space operations in MRI are computationally expensive.
• Latent diffusion models (LDMs) offer potential for reduced computational cost by operating in a lower-dimensional latent space.
• Directly applying LDMs to MRI reconstruction faces challenges including medical fidelity control, domain gap, and latent space data consistency.

Purpose of the Study:

• To introduce a novel Latent Diffusion Prior-based undersampled MRI reconstruction (LDPM) method.
• To address the limitations of existing diffusion models in MRI reconstruction, focusing on computational efficiency and medical accuracy.
• To improve the fidelity and robustness of MRI reconstruction using latent diffusion models.

Main Methods:

• Proposed a novel Latent Diffusion Prior-based undersampled MRI reconstruction (LDPM) framework.
• Implemented a sketch-guided pipeline with a two-step reconstruction strategy for balancing perceptual quality and anatomical fidelity.
• Developed an MRI-optimized Variational Autoencoder (MR-VAE) and a Dual-Stage Sampler for high-fidelity latent space reconstruction.

Main Results:

• The proposed MR-VAE achieved an approximate 3.92 dB PSNR improvement for undersampled MRI reconstruction compared to SD-VAE.
• The LDPM method demonstrated state-of-the-art performance and robustness on the fastMRI dataset.
• Ablation experiments verified the effectiveness of individual modules within the LDPM framework.

Conclusions:

• The LDPM method effectively addresses key challenges in applying latent diffusion models to MRI reconstruction.
• The proposed framework achieves high-fidelity and robust undersampled MRI reconstruction, outperforming existing methods.
• LDPM offers a computationally efficient and accurate solution for accelerated MRI acquisition and reconstruction.