Stabilizing deep tomographic reconstruction: Part A. Hybrid framework and experimental results.

Area of Science:

• Medical imaging
• Deep learning for image reconstruction
• Compressed sensing

Background:

• Popular deep reconstruction networks exhibit instabilities, including image artifacts, missed small features, and performance degradation with increased data.
• These instabilities limit the reliability and accuracy of deep learning-based image reconstruction in scientific applications.

Purpose of the Study:

• To propose and validate a novel framework, Analytic Compressed Iterative Deep (ACID), to address the instabilities in deep reconstruction networks.
• To demonstrate ACID's ability to provide accurate, stable, and robust image reconstruction.

Main Methods:

• Developed the ACID framework, integrating a deep network trained on big data with compressed sensing (CS)-inspired kernel awareness.
• Employed iterative refinement to minimize the data residual relative to real measurements.
• Analyzed the convergence mechanism of ACID iterations under a bounded relative error norm assumption.

Main Results:

• The ACID framework demonstrated accurate and stable image reconstruction.
• ACID successfully eliminated the three reported instabilities: image artifacts, missed features, and performance degradation with increased data.
• ACID showed resilience against adversarial attacks, outperforming classic sparsity-regularized reconstruction methods.

Conclusions:

• The proposed ACID framework effectively stabilizes unstable deep reconstruction networks.
• ACID offers superior performance and robustness compared to existing methods, paving the way for more reliable deep learning in imaging.
• The study provides insights into the convergence properties of iterative deep reconstruction methods.