Ultralow-parameter denoising: Trainable bilateral filter layers in computed tomography.

Area of Science:

• Medical Imaging
• Computational Imaging
• Artificial Intelligence in Healthcare

Background:

• Computed Tomography (CT) is crucial for visualizing 3D structures, offering excellent bone-soft tissue contrast.
• Low-dose CT acquisitions often result in degraded image resolution, necessitating effective denoising techniques.

Purpose of the Study:

• To develop understandable and robust denoising algorithms for CT imaging.
• To minimize radiation dose while preserving data integrity in CT scans.
• To overcome limitations of deep neural networks with excessive parameters in denoising.

Main Methods:

• An open-source CT denoising framework utilizing a data-driven bilateral filter is presented.
• The bilateral filter is optimized via gradient flow, enabling integration into deep learning pipelines.
• Differentiable backprojection layers facilitate denoising across different data domains (raw detector data, reconstructed volume).

Main Results:

• The proposed framework, with only four trainable parameters, rivals state-of-the-art denoising architectures using hundreds of thousands of parameters.
• Competitive denoising performance was demonstrated on x-ray microscope bone data and the 2016 Low Dose CT Grand Challenge dataset.
• Achieved Structural Similarity Index Measures (SSIM) of 0.7094 and 0.9674, and Peak Signal-to-Noise Ratio (PSNR) values of 33.17 and 43.07.

Conclusions:

• The proposed CT denoising pipelines ensure prediction reliability and data integrity due to a minimal number of well-defined trainable parameters.
• This approach offers a robust alternative to conventional deep learning denoising methods, enhancing safety and efficacy in medical imaging.