Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Filter2Noise: Interpretable Self-Supervised Single-Image Denoising for Low-Dose CT with Attention-Guided Bilateral Filtering

Yipeng Sun, Linda-Sophie Schneider, Mingxuan Gu, Siyuan Mei, Chengze Ye, Fabian Wagner, Siming Bayer, Andreas Maier

TL;DR

Low-dose CT denoising is hindered by the scarcity of paired clean data and the need for interpretable denoising. Filter2Noise (F2N) introduces an interpretable self-supervised single-image denoising framework built on an Attention-Guided Bilateral Filter (AGBF) and a downsampling with Euclidean Local Shuffle (ELS) strategy, enabling training from a single noisy image. Key innovations include per-patch spatially varying filter parameters learned via Feature and Sigma Attention, post-training user control over denoising maps, and a multi-scale reconstruction loss with DoG-based regularization that addresses spatial noise correlations, all with a compact ~3.6k parameter footprint. On Mayo Clinic low-dose CT data, F2N achieves up to 4.59 dB PSNR improvements over leading self-supervised methods while offering transparency and region-specific denoising capabilities, suggesting strong potential for clinically interpretable noise reduction across reconstruction kernels.

Abstract

Effective denoising is crucial in low-dose CT to enhance subtle structures and low-contrast lesions while preventing diagnostic errors. Supervised methods struggle with limited paired datasets, and self-supervised approaches often require multiple noisy images and rely on deep networks like U-Net, offering little insight into the denoising mechanism. To address these challenges, we propose an interpretable self-supervised single-image denoising framework -- Filter2Noise (F2N). Our approach introduces an Attention-Guided Bilateral Filter that adapted to each noisy input through a lightweight module that predicts spatially varying filter parameters, which can be visualized and adjusted post-training for user-controlled denoising in specific regions of interest. To enable single-image training, we introduce a novel downsampling shuffle strategy with a new self-supervised loss function that extends the concept of Noise2Noise to a single image and addresses spatially correlated noise. On the Mayo Clinic 2016 low-dose CT dataset, F2N outperforms the leading self-supervised single-image method (ZS-N2N) by 4.59 dB PSNR while improving transparency, user control, and parametric efficiency. These features provide key advantages for medical applications that require precise and interpretable noise reduction. Our code is demonstrated at https://github.com/sypsyp97/Filter2Noise.git .

Filter2Noise: Interpretable Self-Supervised Single-Image Denoising for Low-Dose CT with Attention-Guided Bilateral Filtering

TL;DR

Low-dose CT denoising is hindered by the scarcity of paired clean data and the need for interpretable denoising. Filter2Noise (F2N) introduces an interpretable self-supervised single-image denoising framework built on an Attention-Guided Bilateral Filter (AGBF) and a downsampling with Euclidean Local Shuffle (ELS) strategy, enabling training from a single noisy image. Key innovations include per-patch spatially varying filter parameters learned via Feature and Sigma Attention, post-training user control over denoising maps, and a multi-scale reconstruction loss with DoG-based regularization that addresses spatial noise correlations, all with a compact ~3.6k parameter footprint. On Mayo Clinic low-dose CT data, F2N achieves up to 4.59 dB PSNR improvements over leading self-supervised methods while offering transparency and region-specific denoising capabilities, suggesting strong potential for clinically interpretable noise reduction across reconstruction kernels.

Abstract

Effective denoising is crucial in low-dose CT to enhance subtle structures and low-contrast lesions while preventing diagnostic errors. Supervised methods struggle with limited paired datasets, and self-supervised approaches often require multiple noisy images and rely on deep networks like U-Net, offering little insight into the denoising mechanism. To address these challenges, we propose an interpretable self-supervised single-image denoising framework -- Filter2Noise (F2N). Our approach introduces an Attention-Guided Bilateral Filter that adapted to each noisy input through a lightweight module that predicts spatially varying filter parameters, which can be visualized and adjusted post-training for user-controlled denoising in specific regions of interest. To enable single-image training, we introduce a novel downsampling shuffle strategy with a new self-supervised loss function that extends the concept of Noise2Noise to a single image and addresses spatially correlated noise. On the Mayo Clinic 2016 low-dose CT dataset, F2N outperforms the leading self-supervised single-image method (ZS-N2N) by 4.59 dB PSNR while improving transparency, user control, and parametric efficiency. These features provide key advantages for medical applications that require precise and interpretable noise reduction. Our code is demonstrated at https://github.com/sypsyp97/Filter2Noise.git .

Paper Structure

This paper contains 19 sections, 12 equations, 3 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methodology
  3. Attention-Guided Bilateral Filtering (AGBF)
  4. Feature Attention:
  5. Sigma Attention:
  6. Interpretability and User Control
  7. Training Strategy for Filter2Noise (F2N)
  8. Downsampling
  9. Euclidean Local Shuffle (ELS)
  10. Loss Function
  11. Reconstruction Loss ($L_{rec}$):
  12. Regularization Loss ($L_{reg}$):
  13. Experiments and Discussions
  14. Experimental Configuration
  15. Experimental Results
  16. ...and 4 more sections

Figures (3)

  • Figure 1: (a) The Filter2Noise (F2N) denoising pipeline. (b) The downsampling strategy, following ZS-N2N mansour2023zero. (c) Our proposed Euclidean Local Shuffle (ELS).
  • Figure 2: Denoising Comparison Across Different Noise Conditions.
  • Figure 3: Interpretability Analysis and Ablation Study of F2N-S2.