A Direct Algorithm for Optimization Problems With the Huber Penalty.

Area of Science:

• Medical Imaging and Signal Processing
• Computational Mathematics
• Optimization Algorithms

Background:

• 1-D data fitting is crucial for medical signal processing and image reconstruction.
• Existing iterative methods for penalized data fitting can be computationally intensive.
• The Huber function is effective for penalizing intensity differences in signal processing.

Purpose of the Study:

• To develop a direct, noniterative algorithm for 1-D quadratic data fitting with Huber function penalties.
• To validate the algorithm's performance in medical signal smoothing and sinogram preprocessing.
• To compare the efficiency of the direct algorithm against established iterative methods.

Main Methods:

• A direct algorithm was developed using dynamic programming, reformulating the problem into a sequence of univariate optimization problems.
• The algorithm utilizes back-tracking for efficient solution retrieval.
• Computational aspects and memory costs were analyzed.

Main Results:

• The direct algorithm demonstrated superior efficiency compared to four iterative methods (coordinate descent, Nesterov's accelerated gradient, FISTA, second-order) in 1-D curve smoothing.
• Performance gains increased significantly with the Huber function's curvature.
• Robustness and speed were maintained in sinogram preprocessing, reducing artifacts with minimal computational overhead.

Conclusions:

• The proposed direct algorithm offers a computationally efficient and accurate solution for 1-D penalized data fitting.
• It is highly applicable to medical signal processing tasks like curve smoothing and sinogram preprocessing.
• The algorithm's noniterative nature and superior performance make it a valuable tool for 2-D/3-D image restoration and reconstruction.