Low-dose CT denoising drops the clean-image pairs: align adjacent slices, and cone-beam CT works in image domain

Low-dose CT denoising can be trained without paired clean images, using the supervision signal from two adjacent low-dose slices of the same scan. Supervised methods require aligned noisy-clean image pairs, and such paired datasets are expensive, time-consuming, and sometimes infeasible to acquire. On the Mayo 2016 dataset the method reaches 40.08 dB PSNR, 0.9410 SSIM, and 19.89 HU RMSE, outperforming the compared traditional and self-supervised methods.

An inter-slice matching network estimates the deformation between adjacent slices, and that deformation corrects the denoiser output so the same tissue actually lines up across the two slices. Earlier adjacent-slice methods compute a pixel-wise loss at the same coordinates, but longitudinal anatomical variation makes the same tissue non-rigidly misaligned between slices, so same-coordinate pixels need not belong to the same tissue. The denoiser and matching network are optimized synchronously during training, with reconstruction and correction reinforcing each other; at inference only the denoiser is kept and no spatial resampling is applied.

The correction operation matches pixels that originally sat at different spatial positions and had weaker noise correlation, lowering the noise correlation between corrected adjacent slices. Cone-beam CT has stronger inter-slice noise correlation, and most existing self-supervised cone-beam CT denoising methods rely on projection-domain data. On the real-world cone-beam CT dataset the inter-slice noise correlation falls from 0.2935 before correction to 0.0752 after correction, while a control experiment applying a fixed 0.5-pixel translation changes it only to 0.2917.

The method was trained and tested on the Mayo 2016, Mayo 2020, Piglet, real-world cone-beam CT, and ICASSP cone-beam CT datasets, covering clinical fan-beam CT, single-volume CT, and cone-beam CT. At inference only the denoiser is kept and no spatial resampling is applied, so deployment does not require the matching network. The inter-slice distance parameter is chosen per dataset: 1 for Mayo 2016, Mayo 2020, Piglet, and ICASSP cone-beam CT, and 2 for the real-world cone-beam CT dataset.

The equivalence in Theorem 1 rests on several idealized assumptions that can only be satisfied approximately in practice, because spatial transformations of discrete images involve interpolation and the transition matrix estimated by the matching network may contain errors. The practical training objective should therefore be read as approximately, not strictly, equivalent to supervised training. Increasing the inter-slice distance lowers noise correlation but widens structural differences between slices, making matching harder and producing over-smoothing.