Neighboring Slice Noise2Noise: Self-Supervised Medical Image Denoising from Single Noisy Image Volume

TL;DR

Medical image denoising often requires paired noisy-clean data, which is impractical in many clinical settings. This paper introduces NS-N2N, a self-supervised method that uses neighboring slices within a single noisy volume to form weighted training pairs and trains a denoiser directly in the image domain. The approach employs a weight matrix $W_i$ based on low-pass-filtered residuals and a loss function combining $ ext{L}_{Recon}$, $ ext{L}_{RC}$, and $ ext{L}_{IC}$ to enforce regional and inter-slice consistency, with the total objective $\mathcal{L} = \mathcal{L}_{Recon} + \lambda_{RC}\mathcal{L}_{RC} + \lambda_{IC}\mathcal{L}_{IC}$. Experimental results on synthetic MRI with Rician noise and real low-dose CT show NS-N2N outperforms state-of-the-art single-image self-supervised methods in PSNR/SSIM and achieves favorable processing efficiency, while avoiding device-specific reconstruction issues. Overall, NS-N2N provides a practical, data-efficient, image-domain denoising solution suitable for routine clinical deployment.

Abstract

In the last few years, with the rapid development of deep learning technologies, supervised methods based on convolutional neural networks have greatly enhanced the performance of medical image denoising. However, these methods require large quantities of noisy-clean image pairs for training, which greatly limits their practicality. Although some researchers have attempted to train denoising networks using only single noisy images, existing self-supervised methods, including blind-spot-based and data-splitting-based methods, heavily rely on the assumption that noise is pixel-wise independent. However, this assumption often does not hold in real-world medical images. Therefore, in the field of medical imaging, there remains a lack of simple and practical denoising methods that can achieve high-quality denoising performance using only single noisy images. In this paper, we propose a novel self-supervised medical image denoising method, Neighboring Slice Noise2Noise (NS-N2N). The proposed method utilizes neighboring slices within a single noisy image volume to construct weighted training data, and then trains the denoising network using a self-supervised scheme with regional consistency loss and inter-slice continuity loss. NS-N2N only requires a single noisy image volume obtained from one medical imaging procedure to achieve high-quality denoising of the image volume itself. Extensive experiments demonstrate that the proposed method outperforms state-of-the-art self-supervised denoising methods in both denoising performance and processing efficiency. Furthermore, since NS-N2N operates solely in the image domain, it is free from device-specific issues such as reconstruction geometry, making it easier to apply in various clinical practices.

Figures (5)

• Figure 1: Overview of the proposed NS-N2N method. (a) The proposed weight matrix calculation method. (b)(c) The proposed self-supervised training scheme with regional consistency loss and inter-slice continuity loss.
• Figure 2: Representative image results of different methods in the synthetic experiment, along with their respective error maps compared to the ground truth.
• Figure 3: Comparison of training efficiency between the proposed NS-N2N method and two other dataset-based self-supervised methods, NB2NB and N2V, conducted on the synthetic dataset with noise level $L=7\%$.
• Figure 4: Representative image results of different methods in the real-world experiment.
• Figure 5: Comparison of total time required for denoising an image volume containing 610 slices using different methods.