Learned denoising with simulated and experimental low-dose CT data.

Area of Science:

• Computational imaging
• Machine learning applications
• Medical imaging analysis

Background:

• Machine learning (ML) methods, particularly convolutional neural networks (CNNs), are increasingly used for image processing tasks like noise reduction in computational imaging.
• High-quality training data is crucial for the performance of these ML methods.
• Computed tomography (CT) imaging faces challenges in noise reduction, making ML approaches a focus of research.

Purpose of the Study:

• To comprehensively study the performance differences of ML-based noise reduction algorithms in CT imaging when trained on simulated versus real-world noisy data.
• To compare the effectiveness of two common CNN architectures, U-Net and MSD-Net, for CT image denoising.
• To investigate the impact of training domain (sinogram vs. reconstruction) on denoising performance.

Main Methods:

• Utilized a large 2D computed tomography dataset for machine learning.
• Trained and evaluated U-Net and MSD-Net CNN architectures on both simulated and experimental noisy CT data.
• Compared performance in both the sinogram and reconstruction domains, and explored direct sinogram-to-reconstruction mapping.

Main Results:

• Sinogram denoising showed better performance with simulated data in the sinogram domain, but this did not translate to the reconstruction domain.
• Training on experimental noisy data yielded higher performance for denoising experimental noisy CT data.
• Optimizing algorithms for direct sinogram-to-reconstruction mapping significantly improved model performance.

Conclusions:

• Training ML models on real-world experimental data is superior for CT image denoising compared to simulated data, especially in the reconstruction domain.
• There is a critical need for more advanced noise simulation methods to bridge the gap between simulated and real-world data in CT denoising.
• Matching raw measurement data to high-quality CT reconstructions is vital for effective ML-based denoising.