Conditioning generative latent optimization for sparse-view computed tomography image reconstruction

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Radiology

Background:

• Computed tomography (CT) scans often require numerous X-ray projections, increasing patient radiation dose.
• Acquiring sparse-view CT data leads to significant image degradation with conventional reconstruction methods.
• Existing deep learning methods struggle to generalize to new CT acquisition protocols due to reliance on large, specific training datasets.

Purpose of the Study:

• To develop a novel CT reconstruction method that overcomes limitations of conventional techniques and deep learning models.
• To create a data-independent reconstruction approach that generalizes across different CT acquisition protocols.
• To enable high-quality CT image reconstruction even with minimal or no prior training data.

Main Methods:

• Proposed a conditioned generative latent optimization (cGLO) technique for simultaneous reconstruction of multiple CT slices.
• Evaluated cGLO on full-dose sparse-view CT data with varying numbers of X-ray projections.
• Compared cGLO against Deep Image Prior (DIP) without training data and score-based generative models (SGMs) with training datasets.

Main Results:

• cGLO achieved superior structural SIMilarity (SSIM) compared to SGMs across different dataset sizes.
• The method demonstrated a significant Peak Signal-to-Noise Ratio (PSNR) gain over SGMs, especially with smaller datasets.
• cGLO outperformed Deep Image Prior (DIP), showing a notable PSNR advantage and producing fewer artifacts.

Conclusions:

• The proposed cGLO method offers a robust and efficient solution for sparse-view CT reconstruction.
• cGLO achieves competitive or superior performance compared to state-of-the-art methods.
• This technique is adaptable to various imaging reconstruction tasks, independent of acquisition specifics or data availability.