IEEE Transactions on Medical Imaging 2023
Noise2Sim
Noise Suppression With Similarity-Based Self-Supervised Deep Learning
Current or former lab member. Select a name to see all of their projects.
Abstract
Deep denoising networks usually learn from paired noisy and clean images. Such pairs are rarely available for clinical low-dose CT (LDCT) and photon-counting CT (PCCT). Noise2Sim removes that requirement. It searches the noisy images for sub-images that are similar to each other, and trains the network to map each one to its similar counterparts.
Because the training pairs come from similarity in the data, the method suppresses both independent and correlated noise. Under mild conditions, Noise2Sim is asymptotically equivalent to supervised learning with clean targets.
In experiments on LDCT and PCCT data, Noise2Sim performs as well as or better than supervised methods in visual, quantitative and statistical comparisons.
What’s in the repository
- The
noise2simPython package, installable withpip install noise2sim - Training and test scripts with configurations for LDCT (Mayo and FDA data), PCCT (mouse and chicken leg scans) and natural images (BSD)
- Links to pretrained models and preprocessed datasets
Citation
@article{niu2023noise2sim,
title = {Noise Suppression With Similarity-Based Self-Supervised Deep Learning},
author = {Niu, Chuang and Li, Mengzhou and Fan, Fenglei and Wu, Weiwen and Guo, Xiaodong and Lyu, Qing and Wang, Ge},
journal = {IEEE Transactions on Medical Imaging},
volume = {42},
number = {6},
pages = {1590--1602},
year = {2023},
doi = {10.1109/TMI.2022.3231428}
}