Fourier-Domain Analysis of the Iterative Landweber Algorithm

Area of Science:

• Medical Imaging and Computational Science
• Signal Processing and Inverse Problems

Background:

• Iterative algorithms are crucial for solving inverse problems in medical imaging like X-ray computed tomography (CT), positron emission tomography (PET), and single photon emission computed tomography (SPECT).
• Understanding the convergence properties of these algorithms, such as the Landweber algorithm, is essential for accurate image reconstruction.

Purpose of the Study:

• To analyze the convergence properties of the Landweber algorithm for minimizing quadratic objective functions in two-dimensional (2D) tomography.
• To investigate whether early stopping of the Landweber algorithm yields a minimum-norm solution.
• To determine if low-frequency components consistently converge before high-frequency components in this iterative process.

Main Methods:

• Utilized the frequency-domain transfer function to analyze the behavior of the iterative Landweber algorithm.
• Applied the method to a two-dimensional (2D) tomography problem, relevant to CT, PET, and SPECT imaging.
• Examined the convergence characteristics concerning early stopping and frequency component convergence.

Main Results:

• Demonstrated that early stopping of the Landweber algorithm is not equivalent to obtaining a minimum-norm solution.
• Showed that low-frequency components do not always converge first in the iterative process for 2D tomography.
• The frequency-domain transfer function provided insights into these convergence behaviors.

Conclusions:

• The findings challenge common assumptions about iterative reconstruction in tomography.
• Results indicate that careful consideration of stopping criteria and convergence order is necessary for optimal image reconstruction in CT, PET, and SPECT.
• The frequency-domain analysis offers a valuable tool for understanding iterative algorithm performance.