Practical tradeoffs between noise, quantitation, and number of iterations for maximum likelihood-based...

Area of Science:

• Nuclear Medicine
• Medical Imaging
• Image Reconstruction

Background:

• Filtered backprojection (FBP) is a standard method for emission computed tomography (ECT) image reconstruction.
• Maximum Likelihood-Expectation Maximization (ML) is an iterative reconstruction algorithm with potential for improved image quality.
• Optimizing ML parameters is crucial for balancing noise reduction and quantitative accuracy.

Purpose of the Study:

• To compare ML-based image reconstruction with FBP for ECT.
• To evaluate the impact of different ML reconstruction kernels and iteration counts.
• To assess the effectiveness of sieve filtering and stopping criteria in ML-based reconstruction.

Main Methods:

• Comparison of ML (with Single Pixel and Gaussian kernels, varying iterations, sieve filter) and FBP reconstruction methods.
• Quantitative analysis of image noise and contrast.
• Investigation of feasibility stopping criteria and overrelaxation parameters for ML convergence.

Main Results:

• ML with a Single Pixel kernel offers no quantitative advantage over FBP, except for background noise reduction at low iterations (<50).
• ML with a Gaussian kernel and high iterations (200) requires sieve filtering to mitigate noise and contrast overshoot, yielding minor quantitative benefits over FBP.
• Feasibility stopping criteria control noise but not quantitation errors; overrelaxation accelerates ML convergence without instability.

Conclusions:

• ML-based reconstruction offers limited advantages over FBP in ECT, particularly with specific kernel and iteration choices.
• Sieve filtering is essential for ML Gaussian kernel reconstructions with high iterations to achieve quantitative improvements.
• Parameter optimization, including overrelaxation, is key to efficient and stable ML-based ECT image reconstruction.