Black-box Optimization of CT Acquisition and Reconstruction Parameters: A Reinforcement Learning Approach

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Healthcare

Background:

• Computed Tomography (CT) protocol optimization is essential for balancing image quality and radiation dose.
• Traditional methods require exhaustive parameter testing, which is computationally intensive and impractical.

Purpose of the Study:

• To develop and validate a novel methodology for efficient CT protocol optimization using Virtual Imaging Trials (VITs) and reinforcement learning.
• To demonstrate the accuracy and computational efficiency of the proposed approach compared to exhaustive search methods.

Main Methods:

• Utilized a validated CT simulator and a novel CT reconstruction Toolkit to perform Virtual Imaging Trials on computational phantoms with liver lesions.
• Employed a Proximal Policy Optimization (PPO) agent to optimize parameters including tube voltage, tube current, reconstruction kernel, slice thickness, and pixel size.
• Trained the PPO agent to maximize the Detectability Index (d') for liver lesions in reconstructed CT images.

Main Results:

• The reinforcement learning approach successfully identified the absolute maximum d' for liver lesions across all test cases.
• Achieved optimization with 79.7% fewer steps compared to traditional exhaustive search methods, showcasing significant computational efficiency.
• Demonstrated a flexible framework capable of optimizing for various image quality metrics.

Conclusions:

• Combining VITs and reinforcement learning offers an efficient and robust framework for CT protocol optimization and management.
• The proposed methodology significantly enhances the efficiency of CT protocol development while maintaining or improving diagnostic image quality.
• This AI-driven approach represents a paradigm shift in optimizing medical imaging protocols, paving the way for personalized and precise diagnostics.