Physics-Driven Autoregressive State Space Models for Medical Image Reconstruction

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Science

Background:

• Medical image reconstruction from undersampled data is an ill-posed problem.
• Physics-driven (PD) networks improve reconstruction by combining data consistency with learned priors.
• Current CNNs and transformers struggle with multi-scale contextual structures and non-local dependencies.

Purpose of the Study:

• To introduce MambaRoll, a novel physics-driven autoregressive state space model (SSM) for high-fidelity and efficient medical image reconstruction.
• To address limitations of existing CNNs and transformers in capturing multi-scale dependencies for image reconstruction.
• To improve the disentanglement of artifacts from true anatomical signals in undersampled medical images.

Main Methods:

• Developed MambaRoll, a physics-driven autoregressive state space model (SSM) with an unrolled architecture.
• Employed a hierarchy of scale-specific PD-SSM modules for spatial dependency capture and data consistency enforcement.
• Introduced a Deep Multi-Scale Decoding (DMSD) loss for enhanced scale-aware learning.

Main Results:

• MambaRoll demonstrated improved performance in accelerated MRI and sparse-view CT reconstructions.
• The proposed model effectively handles multi-scale contextual structures and non-local dependencies.
• Achieved high-fidelity and efficient image reconstruction compared to state-of-the-art methods.

Conclusions:

• MambaRoll offers a significant advancement in physics-driven medical image reconstruction.
• The autoregressive SSM architecture with DMSD loss enables superior performance.
• MambaRoll shows promise for clinical applications requiring high-quality imaging from limited data.