Deep learning tomographic reconstruction through hierarchical decomposition of domain transforms

Area of Science:

• Medical Imaging
• Computational Science
• Artificial Intelligence

Background:

• Deep learning (DL) excels at image analysis but struggles with large-scale inverse problems like tomographic reconstruction due to complex integral transforms.
• Existing DL models lack efficiency for non-local and space-variant transforms, limiting practical applications to smaller datasets.
• Prior attempts using fully connected networks were computationally prohibitive for realistic 3D datasets (e.g., 512^6).

Purpose of the Study:

• To develop a novel DL framework for efficient, large-scale tomographic reconstruction.
• To address the limitations of existing DL models in handling complex integral transforms inherent in inverse problems.
• To demonstrate a scalable and data-driven approach for full-size computed tomography (CT) reconstruction.

Main Methods:

• A novel framework casting tomographic reconstruction as a continuum of intermediate representations.
• Breaking down the problem into a sequence of simpler transformations mapped onto an efficient hierarchical network architecture.
• Applying the hierarchical DL approach to computed tomography (CT) image reconstruction with a 512^4 system matrix size.

Main Results:

• The proposed hierarchical DL framework enables efficient reconstruction for large-scale problems, overcoming limitations of previous methods.
• Demonstrated feasibility of full-scale CT reconstruction using a data-driven DL solver, independent of traditional inversion techniques.
• The approach requires exponentially fewer parameters compared to fully connected networks for similar problem sizes.

Conclusions:

• This work introduces a new class of DL solvers for general inverse problems, applicable beyond CT to other imaging modalities.
• The hierarchical DL approach offers potential improvements in signal-to-noise ratio, spatial resolution, and computational efficiency.
• Further development is needed to establish superiority over traditional reconstruction methods, but the feasibility of full-scale learnt reconstruction is shown.