Iterative Residual Optimization Network for Limited-Angle Tomographic Reconstruction.

Area of Science:

• Medical Imaging
• Computational Science
• Artificial Intelligence

Background:

• Limited-angle tomographic reconstruction is an ill-posed inverse problem causing image quality degradation.
• Current deep learning methods often neglect data consistency, leading to suboptimal performance and instability.
• Existing deep reconstruction approaches lack mathematical interpretability.

Purpose of the Study:

• To develop an advanced deep learning model for high-quality limited-angle tomographic reconstruction.
• To address artifacts and improve image fidelity in tomographic reconstruction.
• To provide a mathematically stable and interpretable deep reconstruction framework.

Main Methods:

• Proposed the Iterative Residual Optimization Network (IRON), integrating neural network priors as regularizers.
• Developed a novel optimization objective function to mitigate artifacts from limited-angle data.
• Employed block-coordinate descent for an iterative framework and a convolution-assisted transformer for feature extraction.

Main Results:

• The proposed IRON effectively overcomes false negative and positive artifacts in limited-angle reconstruction.
• The convolution-assisted transformer captures both local and long-range pixel dependencies efficiently.
• IRON demonstrated superior performance compared to state-of-the-art methods on simulated and real cardiac datasets.

Conclusions:

• The Iterative Residual Optimization Network (IRON) offers a robust solution for limited-angle tomographic reconstruction.
• IRON enhances image quality by addressing data consistency and artifact issues.
• The method provides mathematical stability and outperforms existing reconstruction techniques.