Limited parameter denoising for low-dose X-ray computed tomography using deep reinforcement learning.

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Medicine

Background:

• Deep learning (DL) excels in medical imaging but requires extensive data for training deep convolutional networks (CNNs).
• Large DL models may yield unpredictable outcomes due to their high parameter counts.
• Existing DL methods for CT denoising face challenges with data scarcity and potential artifacts.

Purpose of the Study:

• To present a novel CT denoising framework with interpretable behavior.
• To achieve effective CT denoising using limited training data.
• To offer a robust alternative to data-intensive deep learning models.

Main Methods:

• Employed bilateral filtering in projection and volume domains for noise reduction.
• Tuned non-stationary noise parameters (σ) using two deep CNNs trained via Deep-Q reinforcement learning.
• Utilized a custom neural network-based reward function for training the CNNs due to labeling impracticality.

Main Results:

• Achieved significant denoising performance, increasing Peak Signal-to-Noise Ratio (PSNR) from 28.53 to 28.93 and Structural Similarity Index (SSIM) from 0.8952 to 0.9204.
• Outperformed state-of-the-art deep CNNs with significantly fewer parameters (p-value [PSNR] = 0.000, p-value [SSIM] = 0.000).
• Avoided blurring and deep learning artifacts common in other methods; ablation studies confirmed the efficacy of parameter tuning and the reward network.

Conclusions:

• Introduced an interpretable CT denoising framework delivering strong performance, particularly with limited data.
• Demonstrated superior performance compared to current state-of-the-art deep neural networks.
• Future work will focus on accelerating the method and extending its applicability to diverse geometries and anatomical regions.