Noise2Noise Diffusion for Thin-Slice Brain CT Denoising without Clean Training Data.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Image Processing

Background:

• Thin-slice and ultra-high-resolution (UHR) computed tomography (CT) images often exhibit significant noise due to low radiation doses.
• Acquiring clean training data for deep learning-based denoising of thin-slice CT is ethically challenging due to patient radiation exposure concerns.
• Existing supervised learning methods with noise insertion struggle with mismatched noise models and domain shift, impacting performance.

Purpose of the Study:

• To develop a novel deep learning method for high-quality CT image denoising that eliminates the need for clean training data.
• To leverage the strengths of diffusion models and Noise2Noise principles for effective noise reduction in low-dose CT scans.
• To improve quantitative metrics and visual quality of thin-slice CT images compared to existing denoising techniques.

Main Methods:

• Proposed a novel method combining a conditional denoising diffusion probabilistic model (cDDPM) with the Noise2Noise framework.
• Trained the cDDPM to generate a CT slice conditioned on its two adjacent slices, enabling noise independence between input and target.
• Employed multiple sampling and averaging during inference to achieve significant noise reduction in the target CT slice.

Main Results:

• The proposed method demonstrated superior performance in simulated thin-slice brain CT denoising compared to Noise2Noise UNet and supervised DDPM.
• Achieved a lower Mean Absolute Error (MAE) of 2.62 for brain tissues, compared to 4.12 (Noise2Noise UNet) and 3.27 (supervised DDPM).
• Exhibited improved perceptual quality with a lower LPIPS score of 0.0422, versus 0.0917 (Noise2Noise UNet) and 0.0635 (supervised DDPM).

Conclusions:

• The combined diffusion model and Noise2Noise approach effectively denoises thin-slice CT images without requiring clean patient data.
• This method offers a promising solution for improving image quality in low-dose CT, potentially enhancing diagnostic capabilities.
• The technique addresses the limitations of traditional supervised learning in medical image denoising, paving the way for more robust AI applications.