Data Regularization for Streak Artifacts Removal in Self-Supervised Micro-CT Denoising

Abstract

Lens-coupled micro-CT imaging is widely used for its high resolution and noninvasive characteristics. However, due to the low efficiency of its optical system, the reconstructed images suffer from a low signal-to-noise ratio, and it is challenging to acquire sufficient high-quality images. We adapt Noise2Noise for denoising and obtain paired data by dividing a single scan into odd and even projection subsets for separate reconstruction. This process results in undesirable sparse angle artifacts and image structural discrepancies between the noisy pairs. Networks trained on this data tend to mistakenly overfit these discrepancies and introduce streak artifacts during inference. In this article, we propose a self-supervised data regularization-based model that utilizes a unique symmetrical phantom to create pairs of data differing only in noise. This data is input into the network and functions as a regularization mechanism to prevent overfitting. Specifically, we design a fine-tuning strategy based on extremely limited sample data and a mixed datasets training strategy. Both approaches do not need high-quality images. Experimental results show that our method achieves satisfactory denoising effect without introducing artifacts and outperforms the comparison method. This method also generalizes well to unseen samples and various network architectures.