Principal component reconstruction (PCR) for cine CBCT with motion learning from 2D fluoroscopy

Area of Science:

• Medical Imaging
• Image Reconstruction
• Computational Imaging

Background:

• Generating high-temporal-resolution cine CT images is crucial for visualizing dynamic processes in medical imaging.
• Existing methods may face limitations in accurately capturing complex motion patterns.
• Principal Component Reconstruction (PCR) offers a potential framework for low-rank approximation of dynamic imaging data.

Purpose of the Study:

• To develop and validate a novel Principal Component Reconstruction (PCR) technique for generating cine CT images.
• To leverage 2D fluoroscopic training images for motion learning within the PCR framework.
• To achieve high-temporal-resolution 4D image reconstruction using motion-informed priors.

Main Methods:

• The proposed PCR method employs matrix factorization to represent 4D images as spatial principal components and temporal motion coefficients.
• Temporal coefficients are learned from 2D fluoroscopic training projections, synchronized with reconstruction data.
• The reconstruction problem is formulated as a convex optimization problem solved using an alternating direction method of multipliers (ADMM) algorithm, implemented on GPU.

Main Results:

• The PCR method demonstrated feasibility for cine CT reconstruction.
• Reconstruction quality was significantly improved compared to the state-of-the-art PICCS method.
• The GPU-parallelized implementation allows for reconstruction in a few minutes per dataset.

Conclusions:

• The PCR method accurately reconstructs spatial principal components using a priori temporal motion coefficients derived from fluoroscopic training data.
• Cine CT images can be successfully generated as a product of learned temporal motion coefficients and reconstructed spatial principal components.
• This approach offers a robust and efficient solution for high-temporal-resolution dynamic CT imaging.