One network to solve all ROIs: Deep learning CT for any ROI using differentiated backprojection.

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Healthcare

Background:

• Computed tomography (CT) reconstruction for regions of interest (ROI) offers reduced radiation dose and detector size.
• Standard filtered back projection (FBP) methods produce cupping artifacts, while iterative methods are computationally intensive.
• Previous deep learning models struggled with generalizing across different ROI sizes due to image singularities.

Purpose of the Study:

• To develop a generalized deep neural network for CT reconstruction applicable to any ROI size.
• To address the limitations of existing reconstruction methods in handling cupping artifacts and computational complexity.
• To improve the robustness and applicability of deep learning in CT image reconstruction.

Main Methods:

• Designed two types of neural networks: one for ROI size-specific artifacts from FBP images, and another for inverting truncated Hilbert transforms from differentiated backprojection (DBP) data.
• Investigated the generalizability of these networks across varying ROI sizes, pixel sizes, detector pitches, and short-scan starting angles.
• Evaluated performance against existing iterative reconstruction methods.

Main Results:

• The proposed neural networks significantly outperformed existing iterative methods across all tested ROI sizes.
• Achieved substantial reductions in runtime complexity compared to traditional iterative approaches.
• Demonstrated consistent performance improvements across diverse acquisition scenarios, including different detector truncations.

Conclusions:

• The developed neural network approach serves as a generalized CT reconstruction engine suitable for various practical applications.
• Offers superior performance and efficiency compared to existing methods, even with detector truncation.
• Represents a significant advancement in deep learning-based CT image reconstruction.