Learning-Assisted Fast Determination of Regularization Parameter in Constrained Image Reconstruction.

Area of Science:

• Medical Imaging
• Machine Learning
• Computational Science

Background:

• Constrained image reconstruction is crucial in various imaging applications.
• Selecting an optimal regularization parameter is essential for accurate reconstruction but computationally intensive.
• Traditional methods like the L-curve are time-consuming, while modern learning-based methods require extensive training data.

Purpose of the Study:

• To develop a machine learning (ML)-based approach for rapid selection of the optimal regularization parameter in constrained image reconstruction.
• To demonstrate the feasibility and effectiveness of ML in improving reconstruction quality and reducing computational burden.

Main Methods:

• Reconstruction of images using a few pre-selected regularization parameter values.
• Extraction of approximated image quality metrics from initial reconstructions.
• Prediction of true quality metrics using pre-trained neural networks and fusion for optimal parameter determination.

Main Results:

• The proposed ML method significantly reduced the time required for regularization parameter selection compared to L-curve methods.
• Substantially improved image reconstruction quality was achieved.
• The method outperformed state-of-the-art learning-based techniques when trained with limited experimental data.

Conclusions:

• Machine learning provides a feasible and effective solution for determining regularization parameters in constrained image reconstruction.
• The ML approach enhances practical utility by reducing computational load and data requirements.
• This method offers a faster and more accurate alternative for optimizing image reconstruction processes.