Local Pixel Shuffling Delivers More Than 40-Fold Microscopy Background Suppression
Background2Background uses locally shuffled measurements and Noise2Noise-style self-supervision to estimate and remove correlated backgrounds from high-dimensional microscopy data.
Underlying Paper
Background2Background: a self-supervised background removal framework for high dimensional microscopy
High-dimensional microscopy captures features across space, time, and spectrum, but complex backgrounds obscure signals. We present Background2Background (B2B), a self-supervised background removal framework leveraging the backgrounds larger correlation relative to the signal. Local pixel shuffling disrupts signal correlations while preserving background structure, transforming signals into effective noise for Noise2Noise-based background removal. Across fluorescence, stimulated Raman scattering and mid-infrared photothermal imaging, B2B achieves over 40-fold background suppression and reveals hidden signatures.
High-dimensional microscopy can capture spatial, spectral, and temporal information together, but structured backgrounds can obscure weak features. Background2Background (B2B) is a self-supervised framework for background removal that does not require clean-background labels. The authors evaluate it across fluorescence, stimulated Raman scattering (SRS), and mid-infrared photothermal imaging.
Core Contribution
B2B is built on an assumed difference in correlation scale: backgrounds remain correlated over a larger local region than the signals of interest. The method locally shuffles pixels to disrupt signal correlations while preserving background structure. It then uses the shuffled measurements in a Noise2Noise-style training setup to estimate the common background component, which can be subtracted from the original measurement.
This makes the approach a computational correction method rather than a new acquisition modality. Its usefulness depends on the background being more locally correlated than the target signal at the chosen shuffle scale.
Results and Analysis
The paper reports more than 40-fold background suppression across fluorescence, SRS, and mid-infrared photothermal imaging. Its figures show applications across these modalities, including SRS examples and a widefield-fluorescence comparison before and after B2B processing.
Supplementary SRS figures compare B2B with raw SRS, Donut-SRS, rolling-ball, and sliding-paraboloid processing at 2110 cm⁻¹, with corresponding spectral plots. Other supplementary figures show raw and SPEND-processed SRS images at 1048 and 1125 cm⁻¹, a B2B-processed SRS image at 1090 cm⁻¹, and a hyperspectral SRS stack average of U87 cells in the CH region.
The supplementary material also presents an azide signal-extraction procedure based on images at three wavenumbers. In fluorescence, a single-frame widefield image is shown alongside B2B processing after 20 seconds, including magnified views of selected regions.
Caveats in Practice
The reported demonstrations show that B2B can reduce visible background across several microscopy settings, but image-based examples alone do not establish error-free signal recovery. In particular, performance may be sensitive to the local-correlation assumption and the shuffle scale. Smooth target structures or fine-textured backgrounds could make background and signal harder to separate, so the method should be validated for each imaging setting.
Evidence Box
moderateKey Claims
- •Local pixel shuffling disrupts signal correlations while preserving more broadly correlated background structure
- •Noise2Noise-style self-supervision estimates microscopy background without clean-background labels
- •B2B is demonstrated for fluorescence, SRS, and mid-infrared photothermal imaging
Key Results
- •More than 40× background suppression is reported across the three imaging modalities
- •Supplementary SRS comparisons include B2B, Donut-SRS, rolling-ball, and sliding-paraboloid processing
- •Supplementary figures include raw, SPEND-processed, and B2B-processed SRS examples
- •A single-frame widefield-fluorescence comparison shows B2B processing after 20 seconds
Limitations & Caveats
- •The method relies on background being more locally correlated than the target signal
- •Visual comparisons alone do not establish quantitative signal-recovery accuracy
- •Performance should be validated when target structures are smooth or backgrounds contain fine texture
