Low-dose 4DCT reconstruction via temporal nonlocal means

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Computational Imaging

Background:

• Four-dimensional computed tomography (4DCT) is crucial for cancer radiotherapy, enabling accurate tumor delineation and motion assessment.
• However, 4DCT's prolonged scanning increases radiation dose, posing a clinical challenge.
• Reducing radiation dose in 4DCT is essential for patient safety and frequent monitoring.

Purpose of the Study:

• To develop a novel algorithm for reconstructing 4DCT images from undersampled projections acquired at low milliampere-second (mAs) levels.
• The primary goal is to significantly reduce the imaging radiation dose associated with 4DCT scans.
• To maintain or improve image quality despite dose reduction and undersampling.

Main Methods:

• A Temporal Nonlocal Means (TNLM) method was developed to leverage inter-phase similarities in 4DCT data.
• A simultaneous reconstruction approach minimized a cost function including data fidelity and TNLM regularization terms.
• The reconstruction problem was solved using modified forward-backward splitting and Gauss-Jacobi iteration, accelerated by GPU implementation.

Main Results:

• The TNLM algorithm produced superior 4DCT images with reduced noise and streaking artifacts compared to standard Filtered Backprojection (FBP).
• Quantitative improvements included contrast-to-noise ratio increases of 3.9-10.2x and signal-to-noise ratio increases of 2.1-5.9x.
• The algorithm effectively suppressed streaking artifacts in undersampled data and reconstructed ten phases within 40-90 seconds on a GPU.

Conclusions:

• The developed TNLM algorithm significantly outperforms conventional FBP in reducing artifacts from undersampling and noise from low-dose acquisitions.
• This novel approach enables high-quality 4DCT imaging at substantially reduced radiation doses.
• The TNLM algorithm shows promise for improving the clinical applicability of 4DCT in radiotherapy.