Abstract
Ultra-low-dose computed tomography (ULDCT) can substantially reduce radiation exposure, but severe photon starvation and quantum noise introduce structured streak artifacts that obscure low-contrast anatomy. Image-domain deep denoisers can suppress random noise, yet they often over-smooth subtle textures or produce outputs that are inconsistent with the measured projection data. We propose a physics-guided Noise2Noise plug-and-play iterative reconstruction framework for ULDCT, where a learned denoising prior is embedded within an ADMM reconstruction loop. Each iteration alternates between a projection domain data-consistency update and an image-domain denoising step, combining CT acquisition physics with learned anatomical priors. The denoiser uses self-supervised criss-cross Noise2Noise pre-training on independently corrupted low-dose reconstructions, followed by supervised calibration with texture- and contrast-aware losses. This hybrid training provides a noise-aware initialization while improving anatomical detail preservation. Experiments on simulated multi-dose CT data and de-identified clinical low-dose scans show that the proposed method enables high-quality reconstruction from approximately 75–88% lower tube-current settings relative to the high-dose reference, while suppressing structured and stochastic noise more effectively than filtered backprojection, standalone denoisers, and supervised-prior PnP reconstruction. On clinical scans, the proposed method achieves the highest contrast-to-noise ratio among all evaluated methods, indicating improved soft-tissue visibility under realistic low-dose conditions. -2mm Ultra-low-dose CT Plug-and-play reconstruction Iterative reconstruction Noise2Noise ADMM Deep denoising
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/RIME_016.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=QlUEh0HLCF
BibTex
@InProceedings{DutSay_PhysicsGuided_MICCAISAT2026,
author = { Dutta, Sayantan AND Chatterjee, Sudhanya AND Galande, Ashwini AND Shriram, K. S. AND Das, Bipul},
title = { { Physics-Guided Noise2Noise Plug-and-Play MBIR for Ultra-Low-Dose CT Reconstruction } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17277},
month = {pending},
page = {pending}
}