Denoising and deconvolving CT images of unknown origin: comparing linear Wiener-deconvolution with deep convolutional neural network Noise2Inverse
Low-dose CT scans are very fast, but they feature strong pixel noise which, in turn, invites for efficient denoising the same way that strong image blur invites for sharpening the images. Wiener-deconvolution serves as a starting point, combining these two operations in a linear Fourier filter. It...
Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Article |
| Language: | deu |
| Published: |
NDT.net
2025-02-01
|
| Series: | e-Journal of Nondestructive Testing |
| Online Access: | https://www.ndt.net/search/docs.php3?id=30718 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1832086559649693696 |
|---|---|
| author | Simon Zabler Antoine Klos Pierre Lhuissier Luc Salvo Maziyar Farahmandi Simon Wittl |
| author_facet | Simon Zabler Antoine Klos Pierre Lhuissier Luc Salvo Maziyar Farahmandi Simon Wittl |
| author_sort | Simon Zabler |
| collection | DOAJ |
| description |
Low-dose CT scans are very fast, but they feature strong pixel noise which, in turn, invites for efficient denoising the same way that strong image blur invites for sharpening the images. Wiener-deconvolution serves as a starting point, combining these two operations in a linear Fourier filter. It may apply in 2D or 3D and requires merely estimates of the Signal-to-Noise Ratio SNR2 as well as the system Modulation Transfer Function (MTF). Meanwhile, any linear filter is unable to adapt to local contrast variations in the images. Therefore, convolutional neural networks (CNN) such as MSDnet or UNet can significantly improve reconstructions from low-dose CT scans thanks to their non-linear nature. The Noise2Inverse framework allows for training these CNN using a split-and-merge of two subsets generated from one raw CT dataset. This study investigates denoising by N2I with respect to the input data quality, the CNN network architecture and the fidelity of the features in the resulting denoised volume images. Unlike Wiener filtering, N2I is far more effective in removing pixel noise. Yet, these models give the vague impression that they might generate ghosts of small features (pores or spots) in the object. Image quality metrics like Fourier Shell Correlation and SNR2 power spectra show clear improvement. In particular, Noise2Inverse preserves the signal power spectrum of all scans reliably. Nevertheless, locally inspecting image features must remain a core technique when the CNN models are applied, i.e. for industrial inspection.
|
| format | Article |
| id | doaj-art-1c58bf46b7bc4bb9b778bbbc36a216a8 |
| institution | Kabale University |
| issn | 1435-4934 |
| language | deu |
| publishDate | 2025-02-01 |
| publisher | NDT.net |
| record_format | Article |
| series | e-Journal of Nondestructive Testing |
| spelling | doaj-art-1c58bf46b7bc4bb9b778bbbc36a216a82025-02-06T10:48:18ZdeuNDT.nete-Journal of Nondestructive Testing1435-49342025-02-0130210.58286/30718Denoising and deconvolving CT images of unknown origin: comparing linear Wiener-deconvolution with deep convolutional neural network Noise2InverseSimon ZablerAntoine KlosPierre LhuissierLuc SalvoMaziyar FarahmandiSimon Wittlhttps://orcid.org/0009-0006-0365-7018 Low-dose CT scans are very fast, but they feature strong pixel noise which, in turn, invites for efficient denoising the same way that strong image blur invites for sharpening the images. Wiener-deconvolution serves as a starting point, combining these two operations in a linear Fourier filter. It may apply in 2D or 3D and requires merely estimates of the Signal-to-Noise Ratio SNR2 as well as the system Modulation Transfer Function (MTF). Meanwhile, any linear filter is unable to adapt to local contrast variations in the images. Therefore, convolutional neural networks (CNN) such as MSDnet or UNet can significantly improve reconstructions from low-dose CT scans thanks to their non-linear nature. The Noise2Inverse framework allows for training these CNN using a split-and-merge of two subsets generated from one raw CT dataset. This study investigates denoising by N2I with respect to the input data quality, the CNN network architecture and the fidelity of the features in the resulting denoised volume images. Unlike Wiener filtering, N2I is far more effective in removing pixel noise. Yet, these models give the vague impression that they might generate ghosts of small features (pores or spots) in the object. Image quality metrics like Fourier Shell Correlation and SNR2 power spectra show clear improvement. In particular, Noise2Inverse preserves the signal power spectrum of all scans reliably. Nevertheless, locally inspecting image features must remain a core technique when the CNN models are applied, i.e. for industrial inspection. https://www.ndt.net/search/docs.php3?id=30718 |
| spellingShingle | Simon Zabler Antoine Klos Pierre Lhuissier Luc Salvo Maziyar Farahmandi Simon Wittl Denoising and deconvolving CT images of unknown origin: comparing linear Wiener-deconvolution with deep convolutional neural network Noise2Inverse e-Journal of Nondestructive Testing |
| title | Denoising and deconvolving CT images of unknown origin: comparing linear Wiener-deconvolution with deep convolutional neural network Noise2Inverse |
| title_full | Denoising and deconvolving CT images of unknown origin: comparing linear Wiener-deconvolution with deep convolutional neural network Noise2Inverse |
| title_fullStr | Denoising and deconvolving CT images of unknown origin: comparing linear Wiener-deconvolution with deep convolutional neural network Noise2Inverse |
| title_full_unstemmed | Denoising and deconvolving CT images of unknown origin: comparing linear Wiener-deconvolution with deep convolutional neural network Noise2Inverse |
| title_short | Denoising and deconvolving CT images of unknown origin: comparing linear Wiener-deconvolution with deep convolutional neural network Noise2Inverse |
| title_sort | denoising and deconvolving ct images of unknown origin comparing linear wiener deconvolution with deep convolutional neural network noise2inverse |
| url | https://www.ndt.net/search/docs.php3?id=30718 |
| work_keys_str_mv | AT simonzabler denoisinganddeconvolvingctimagesofunknownorigincomparinglinearwienerdeconvolutionwithdeepconvolutionalneuralnetworknoise2inverse AT antoineklos denoisinganddeconvolvingctimagesofunknownorigincomparinglinearwienerdeconvolutionwithdeepconvolutionalneuralnetworknoise2inverse AT pierrelhuissier denoisinganddeconvolvingctimagesofunknownorigincomparinglinearwienerdeconvolutionwithdeepconvolutionalneuralnetworknoise2inverse AT lucsalvo denoisinganddeconvolvingctimagesofunknownorigincomparinglinearwienerdeconvolutionwithdeepconvolutionalneuralnetworknoise2inverse AT maziyarfarahmandi denoisinganddeconvolvingctimagesofunknownorigincomparinglinearwienerdeconvolutionwithdeepconvolutionalneuralnetworknoise2inverse AT simonwittl denoisinganddeconvolvingctimagesofunknownorigincomparinglinearwienerdeconvolutionwithdeepconvolutionalneuralnetworknoise2inverse |