Precision and accuracy of regional radioactivity quantitation using the maximum likelihood EM reconstruction...

Area of Science:

• Medical Imaging
• Nuclear Medicine
• Computational Science

Background:

• Maximum Likelihood (ML) reconstruction using the Expectation-Maximization (EM) algorithm in emission tomography is well-studied for imaging characteristics.
• Fewer studies have focused on the precision and accuracy of ML estimates for regional radioactivity concentration.

Purpose of the Study:

• To evaluate the bias and variability of ML estimates for regional radioactivity concentration.
• To compare ML reconstruction with Filtered Backprojection (FBP) using a realistic brain phantom.
• To assess the impact of pixel size, smoothing, and region size reduction on ML and FBP reconstructions.

Main Methods:

• Developed a realistic brain slice simulation from segmented MRI data (gray matter, white matter, CSF).
• Generated PET sinogram data incorporating detector resolution, efficiencies, attenuation, scatter, and randoms.
• Performed ML and FBP reconstructions on noisy data at various count levels and analyzed ROI values.

Main Results:

• ML reconstructions (3,000 iterations) for 1-cm² gray matter ROIs showed a -6%+/-2% bias, reducible to 0%+/-3% by excluding the outer 1-mm rim.
• FBP reconstructions on full-size ROIs exhibited a -15%+/-4% bias but with 50% less noise than ML.
• Shrinking FBP regions partially compensated for bias, increasing noise to ML levels; ML image smoothing yielded bias comparable to FBP with slightly less noise.

Conclusions:

• ML reconstruction provides more accurate regional radioactivity concentration estimates in PET, particularly when minimal bias is critical.
• Despite higher computational demands, ML is advantageous for applications requiring high precision in quantitative analysis.
• Image processing techniques like ROI rim exclusion and region size adjustment can further optimize quantitative accuracy for both ML and FBP methods.