MEMORY-EFFICIENT DEEP END-TO-END POSTERIOR NETWORK (DEEPEN) FOR INVERSE PROBLEMS

Area of Science:

• Medical Imaging
• Computational Neuroscience
• Machine Learning

Background:

• End-to-End (E2E) unrolled optimization frameworks are promising for Magnetic Resonance (MR) image recovery.
• These deterministic methods face challenges with high memory usage during training and lack posterior distribution sampling capabilities.

Purpose of the Study:

• To introduce a memory-efficient approach for E2E learning of the posterior distribution in MR image reconstruction.
• To enable uncertainty quantification alongside image recovery.

Main Methods:

• A novel framework combining a data-consistency likelihood term and a CNN-parameterized prior energy model.
• E2E learning of CNN weights via maximum likelihood optimization.
• Maximum A Posteriori (MAP) optimization for image recovery from undersampled MR data.

Main Results:

• The proposed method achieves comparable performance to memory-intensive E2E unrolled algorithms.
• It outperforms existing memory-efficient counterparts in MR image reconstruction.
• The framework successfully generates uncertainty maps derived from posterior distribution sampling.

Conclusions:

• This memory-efficient E2E learning framework advances MR image reconstruction.
• It offers a viable solution for high-dimensional (3D+) MR imaging.
• The ability to sample the posterior distribution provides valuable uncertainty information.