Learning low-dose CT degradation from unpaired data with flow-based model.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Growing interest in low-dose computed tomography (LDCT) to reduce patient radiation exposure.
• LDCT images suffer from complex noise, hindering diagnostic accuracy.
• Supervised deep learning methods for LDCT denoising require large paired datasets, which are clinically difficult to obtain.

Purpose of the Study:

• To develop a method for LDCT denoising that does not require paired training data.
• To train a neural network in a weakly supervised manner to simulate LDCT images from normal-dose CT (NDCT) images.
• To utilize simulated training pairs for supervised deep denoising networks.

Main Methods:

• Proposed a weakly supervised, flow-based model to learn LDCT degradation from unpaired LDCT and NDCT images.
• Modeled image degradation in a shared latent space by projecting both LDCT and NDCT images.
• Trained the model using negative log-likelihood loss, eliminating the need for paired data; generated LDCT images from NDCT images for subsequent supervised denoising network training.

Main Results:

• Achieved superior LDCT image simulation performance compared to CycleGAN, validated by radial noise power spectrum analysis.
• Generated image pairs effectively trained deep learning denoising networks (REDCNN, TransCT).
• Demonstrated superior LDCT denoising performance on abdomen and chest CT datasets, outperforming traditional methods, CycleGAN-trained networks, and transfer learning methods, with visual results comparable to supervised networks.

Conclusions:

• Developed a flow-based method for LDCT degradation learning using unpaired data, enabling impressive LDCT synthesis.
• Generated paired data can train neural networks for effective LDCT denoising.
• The proposed method yields denoising results superior to traditional and weakly supervised approaches, and comparable to fully supervised deep learning methods.