Axial Self-Similarity based modular Noise2Noise for progressive LDCT denoising without ground truth.

Low-dose computed tomography (LDCT) has emerged as an alternative to standard-dose computed tomography (SDCT) to reduce radiation exposure. However, this dose reduction comes at the cost of increased noise and artifacts, which can compromise diagnostic accuracy. Although deep learning–based denoising methods have shown promising performance in LDCT denoising, their reliance on paired noisy-clean or noisy-noisy datasets limits practical applicability. Additionally, deep denoisers trained via end-to-end mapping on paired data typically produce a single denoised output with a fixed noise reduction level. In clinical practice, LDCT is often acquired at variable dose levels, and fixed-level denoising may not adapt well to such variability. In this study, we present a modular self-supervised denoising method by building upon the Noise2Noise principle. Axial self-similarity among neighboring slices within a CT volume has been exploited to relax the need for paired noisy-clean or noisy-noisy data. In particular, a self-supervised loss that incorporates inter-slice congruence and neighboring slice consistency is developed to enable effective learning directly from a single LDCT scan. Furthermore, an end-to-process training scheme is employed through the modular cascading of multiple identical denoising module. Each cascade refines the output of the previous stage, resulting in denoised outputs at multiple noise reduction levels across stages. Experiments on two publicly available LDCT datasets demonstrate the effectiveness of our method in reducing noise while preserving structural details, highlighting its clinical potential.