Exploring the Influence of Initial Estimates on Iterative Maximum Likelihood Expectation-maximization in Tomographic...

Area of Science:

• Medical Imaging
• Computational Imaging
• Image Reconstruction

Background:

• Iterative reconstruction algorithms like Maximum Likelihood Expectation Maximization (MLEM) are crucial for tomographic imaging.
• The choice of initial image estimate can influence the convergence and quality of MLEM reconstructions.
• Filtered Back Projection (FBP) is a common, though often suboptimal, method for initial image estimation.

Purpose of the Study:

• To investigate the impact of different initial image estimates on MLEM tomographic reconstruction.
• To compare the convergence patterns of MLEM using uniform (ones, zeros) and FBP initial images.
• To analyze the Mean Squared Error (MSE) curves for each initial estimate during reconstruction.

Main Methods:

• Two digital phantoms were used: a noiseless object and a noise-degraded lower thorax section.
• Sinograms were generated via the radon transform, with noise added to the thorax phantom's sinogram.
• MLEM reconstruction was performed using three initial estimates: uniform 'ones', uniform 'zeros', and an FBP image, with varying iteration counts.

Main Results:

• Uniform 'ones' and 'zeros' initial estimates produced nearly identical convergence curves.
• The FBP initial estimate showed faster convergence towards the true object in early iterations.
• Normalized MSE decreased sharply within the first 5-10 iterations, plateauing thereafter for all estimates.

Conclusions:

• Uniform initial images (ones or zeros) are effective and comparable to FBP for MLEM reconstruction.
• The choice of initial estimate has a minimal impact on the overall convergence pattern.
• Uniform initial estimates are a practical and potentially time-saving alternative to FBP.