Subspace Model-Assisted Deep Learning for Improved Image Reconstruction.

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Medicine

Background:

• Image reconstruction from limited data is an ill-posed problem requiring a priori information.
• Traditional methods use general image properties (sparsity, low-rank).
• Deep learning (DL) offers improved reconstruction but requires extensive data, often unavailable in medical imaging, leading to sensitivity and generalization issues.

Purpose of the Study:

• To address the sensitivity to data perturbations and limited generalization of DL-based image reconstruction.
• To propose a novel method synergistically integrating model-based and data-driven learning.
• To enhance the practical utility of DL-based image reconstruction in medical applications.

Main Methods:

• A three-component method combining linear vector space for global features, a deep network for manifold mapping, and an unrolling-based network for local residuals with sparsity modeling.
• Utilizing magnetic resonance imaging (MRI) data for evaluation.

Main Results:

• Demonstrated improved image reconstruction quality.
• Showcased enhanced performance in the presence of data perturbations.
• Validated improved generalization capabilities, particularly with novel image features.

Conclusions:

• The proposed method effectively integrates model-based and data-driven learning to overcome limitations of current DL reconstruction techniques.
• This synergistic approach enhances robustness and generalization, showing promise for practical medical imaging applications.
• The method offers a pathway to more reliable and versatile DL-based image reconstruction.