Intelligent Parameter Tuning in Optimization-Based Iterative CT Reconstruction via Deep Reinforcement Learning.

Area of Science:

• Computer Vision
• Medical Imaging
• Machine Learning

Background:

• Image processing often involves complex optimization problems with multiple parameters.
• Manual tuning of these parameters is critical for solution quality but can be tedious and impractical.
• Existing methods lack efficient automated parameter adjustment strategies.

Purpose of the Study:

• To develop an automated system for parameter tuning in image processing using deep reinforcement learning.
• To mimic human-like intuition in adjusting parameters for optimization problems.
• To improve the efficiency and effectiveness of parameter tuning in computed tomography (CT) reconstruction.

Main Methods:

• Formulated image processing as an optimization problem with a multi-term objective function.
• Employed deep reinforcement learning to train a parameter-tuning policy network (PTPN).
• PTPN maps image patches to parameter adjustment directions and amplitudes for end-to-end training.

Main Results:

• Successfully trained a PTPN for automated parameter adjustment in optimization-based iterative CT reconstruction.
• Demonstrated that the PTPN-guided reconstruction yields CT images of comparable or superior quality to manual tuning.
• The approach effectively handles pixel-wise total-variation regularization terms.

Conclusions:

• Deep reinforcement learning offers a viable solution for automated parameter tuning in image processing.
• The proposed PTPN approach automates a critical yet challenging aspect of optimization-based image reconstruction.
• This method has the potential to enhance the quality and efficiency of medical image reconstruction and other image processing tasks.