Fast iterative reconstruction for photoacoustic tomography using learned physical model: Theoretical validation.

Area of Science:

• Medical Imaging
• Computational Science
• Artificial Intelligence in Healthcare

Background:

• Iterative reconstruction algorithms excel in challenging medical imaging scenarios (compressed, sparse, limited-view).
• Traditional iterative methods are computationally intensive, requiring many iterations and manual parameter tuning for optimal performance.
• Learned iterative reconstruction (e.g., model-based learning) speeds up convergence but still involves repeated physical model evaluations.

Purpose of the Study:

• To develop a Fast Iterative Reconstruction (FIRe) algorithm.
• To integrate a learned physical model within a learned iterative reconstruction framework.
• To significantly reduce reconstruction time while preserving robust image quality.

Main Methods:

• Developed the Fast Iterative Reconstruction (FIRe) algorithm incorporating a learned physical model.
• Proposed an efficient recursive training scheme to mitigate high memory demands of learned iterative reconstruction.
• Evaluated FIRe against traditional and learned iterative reconstruction methods.

Main Results:

• FIRe achieved reconstruction performance comparable to learned iterative reconstruction methods.
• Demonstrated a 9x reduction in computation time compared to learned iterative reconstruction.
• Achieved a 620x reduction in computation time compared to variational reconstruction methods.

Conclusions:

• The proposed FIRe algorithm offers a significant computational advantage for iterative reconstruction in medical imaging.
• FIRe's efficient training scheme addresses memory limitations, making advanced reconstruction more accessible.
• This approach enables faster, high-quality medical image reconstruction without compromising performance.