AirNet: Fused analytical and iterative reconstruction with deep neural network regularization for sparse-data CT.

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Healthcare

Background:

• Sparse-data computed tomography (CT) is prevalent in applications like breast tomosynthesis and C-arm CT.
• Reconstructing images from highly undersampled data in sparse-data CT presents significant challenges.

Purpose of the Study:

• To develop a novel data-driven image reconstruction method for sparse-data CT using deep neural networks (DNNs).
• To improve image quality and clinical utility in scenarios with limited CT data.

Main Methods:

• Developed AirNet, a DNN-based method integrating analytical reconstruction (AR), iterative reconstruction (IR), and DNNs.
• Utilized a fused analytical and iterative reconstruction (AIR) framework with modified proximal forward-backward splitting (PFBS).
• Unrolled PFBS into iterative updates, incorporating AR (e.g., FBP) for data fidelity and DNNs with residual learning for regularization.

Main Results:

• AirNet demonstrated the fastest convergence rate in validation loss compared to other methods.
• The method showed robustness to noise and variations between training and test data.
• AirNet achieved superior image reconstruction quality and improved radiotherapy treatment planning (photon and proton therapy) compared to state-of-the-art methods.

Conclusions:

• AirNet represents a significant advancement in sparse-data CT image reconstruction.
• The method offers superior visual and quantitative image reconstruction across various sparse-data scenarios.
• AirNet provides improved treatment plan quality for both photon and proton therapy using sparse-data CT.