Quantitative comparison of FBP, EM, and Bayesian reconstruction algorithms for the IndyPET scanner

Area of Science:

• Medical Imaging
• Nuclear Medicine
• Computational Imaging

Background:

• Positron Emission Tomography (PET) imaging relies on reconstruction algorithms to create images from detected photons.
• Traditional algorithms like Filtered Backprojection (FBP) have limitations in resolution and noise handling.
• Iterative algorithms, such as Expectation-Maximization (EM) and Bayesian methods, offer potential for improved image quality.

Purpose of the Study:

• To quantitatively compare FBP, EM, and Bayesian reconstruction algorithms for the IndyPET scanner.
• To evaluate the impact of incorporating an empirical system kernel for resolution recovery.
• To assess reconstruction quality using both phantom and simulated data.

Main Methods:

• Utilized an empirical system kernel derived from line source phantom scans for EM and Bayesian algorithms.
• Applied FBP, EM, and Bayesian reconstruction to bar phantom, Hoffman brain phantom, and simulated lesion data.
• Quantitatively analyzed reconstruction quality using bias-variance and mean square error metrics.

Main Results:

• Without empirical kernel, FBP, EM, and Bayesian algorithms showed similar performance.
• Inclusion of the empirical kernel led to superior reconstructions with iterative algorithms over FBP.
• Bayesian methods demonstrated better performance than EM when using the empirical kernel.

Conclusions:

• Accurate system response modeling, via empirical kernels, is crucial for PET image reconstruction.
• Bayesian reconstruction algorithms, combined with accurate system models, yield significant improvements in PET image quality.
• This approach enhances resolution recovery and quantitative accuracy in small and intermediate field-of-view PET imaging.