4D-AirNet: a temporally-resolved CBCT slice reconstruction method synergizing analytical and iterative method with...

Area of Science:

• Medical Imaging
• Image Reconstruction
• Deep Learning Applications

Background:

• Four-dimensional (4D) cone-beam CT (CBCT) reconstructs dynamic 3D volumes from limited projection data, posing a significant sparse-data challenge.
• High-quality reconstruction of temporally-resolved phases in 4D CBCT is crucial for accurate medical imaging, particularly in applications like lung imaging.

Purpose of the Study:

• To develop a novel deep learning-based method, 4D-AirNet, for high-quality, temporally-resolved 4D CBCT slice reconstruction.
• To synergize analytical and iterative reconstruction (AIR) techniques with deep learning (DL) to address the sparse-data problem in 4D CBCT.

Main Methods:

• Introduced 4D-AirNet, an unrolling method based on the proximal forward-backward splitting (PFBS) optimization framework for fused AIR.
• Investigated three strategies: random-phase (RP), prior-guided (PG), and all-phase (AP) training and reconstruction, with AP-AirNet training all phases simultaneously.
• Incorporated dense connectivity in networks and explored joint regularization of DL with spatiotemporal total variation (TV).

Main Results:

• 4D-AirNet methods demonstrated superior performance compared to conventional iterative (TV) and deep learning (LEARN) methods in simulated sparse-data 4D CBCT scans.
• The all-phase (AP) strategy within 4D-AirNet yielded the best overall reconstruction quality among the evaluated methods.
• Dense connectivity improved reconstruction quality, highlighting its benefit in the 4D-AirNet architecture.

Conclusions:

• The proposed 4D-AirNet method effectively enhances 4D CBCT image reconstruction quality, particularly under sparse-data conditions.
• The all-phase (AP) approach within 4D-AirNet shows significant promise for achieving superior temporal resolution and image fidelity in dynamic CBCT.