Noise2Noise diffusion for thin-slice brain CT denoising without clean training data

Thin-slice and ultra-high-resolution (UHR) computed tomography (CT) images usually suffer from excessive noise due to limited radiation dose being distributed into small detector units. In the context of deep learning-based image denoising, it is challenging to obtain clean training data from real patients for thin-slice CT because it is unethical to apply excessively high doses to the patients. Supervised learning with noise insertion would face unmatched noise models and possibly domain shift of the training data, which leads to deteriorated model performance. In this work, we proposed a novel method that combined the diffusion model with Noise2Noise, which achieved high-quality noise reduction without requiring clean training data. A conditional denoising diffusion probabilistic model (cDDPM) was trained to sample a CT slice from its two adjacent slices. Because of the noise independence between the input and target, DDPM would sample another noise realization of the target slice. During the inference, the trained DDPM was sampled multiple times to acquire multiple samples of the target slice, which were averaged for a slice with lower noise. The method was validated with simulated thin-slice brain CTs, demonstrating improved quantitative metrics and visual impressions compared to Noise2Noise UNet and supervised DDPM with a slightly mismatched noise model. The mean absolute errors (MAE) of the brain tissues were 4.12, 3.27, and 2.62 for Noise2Noise UNet, supervised DDPM, and the proposed method, respectively. The perceptual loss (LPIPS) was 0.0917, 0.0635, and 0.0422 for the three methods, respectively.