Fast Variance Prediction for Iteratively Reconstructed CT Images With Locally Quadratic Regularization

Area of Science:

• Medical Imaging
• Computational Imaging
• Image Reconstruction

Background:

• Accurate noise prediction in CT images is crucial for evaluating reconstruction algorithms.
• Existing methods for noise power spectrum (NPS) prediction are computationally intensive or limited in scope.
• Local noise variance and NPS are key parameters for image analysis and optimization.

Purpose of the Study:

• To develop a computationally efficient method for predicting reconstructed image variance and local NPS.
• To enable fast noise analysis for statistical iterative reconstruction methods with quadratic regularization.
• To provide a tool for improving CT image quality and feature detectability.

Main Methods:

• A novel method utilizing only projection data for noise prediction was developed.
• The approach is designed for locally shift-invariant CT geometries with adequate angular sampling.
• Computation time is comparable to a single back-projection, significantly faster than previous techniques.

Main Results:

• The method successfully generates a variance map for the entire image.
• It achieves reasonable accuracy, particularly for regions away from image edges.
• Validation was performed using both simulated and real thorax phantom CT data.

Conclusions:

• The proposed method offers a fast and practical approach to predict noise properties in iteratively reconstructed CT images.
• It overcomes the computational limitations of prior methods, making noise analysis more accessible.
• Further refinement may be needed for accurate predictions near edges due to regularization effects.