Ziel/Aim The image quality achieved in iterative PET image reconstruction is influenced by
several internal and user-settable parameters (number of iterations and subsets, PSF
model, etc.). Typically, there are more than 3 user-settable parameters involved,
interacting in a non-intuitive way. Reasonable settings typically are obtained interactively
by try-and-error which is highly subjective. This proof-of-concept work proposes a
method for automated reconstruction parameters optimization for a given, preselected
image quality metric.
Methodik/Methods In out approach, we reconstruct images of cylindrical phantom with six “hot” sphere
inserts simulating lesions of different sizes and target-to-background activity concentration
ratios (20:1, 10:1, 5:1). 4 parameters of our in-house reconstruction tool THOR [1] were varied during optimization: no. of iterations and subsets, tube of response
(ToR) radius, Gaussian post filter FWHM. As image quality metric we chose the weighted
sum of standard deviation of contrast recovery coefficients of all 6 inserts (as a
surrogate for image resolution), the image noise, and Gibbs artifacts. This metric
is minimized with Bayesian optimization method using Gaussian process as a surrogate
function. The reconstruction parameters resulting in the minimum metric value were
chosen.
Ergebnisse/Results The optimization process lasted for 50 iterations. The resulting reconstruction parameters
were: no. of iterations/subsets = 2/21, ToR radius = 2.95mm, Gaussian filter FWHM = 4.0mm.
The resulting images show [4.7-5.8]mm resolution and 14 % noise level. Gibbs artifacts
level was found to be below 3.5 %.
Schlussfolgerungen/Conclusions Our framework for reconstruction protocol optimization is capable of deriving reasonable
reconstruction parameters in a fully automated manner. The presented approach might
also be used to improve and objectify the comparison of different image reconstruction
algorithms.
