Fig. 2: Benchmark of LF-denoising with various denoising methods on the data captured by LFM. | Nature Communications

Fig. 2: Benchmark of LF-denoising with various denoising methods on the data captured by LFM.

From: Leveraging spatial-angular redundancy for self-supervised denoising of 3D fluorescence imaging without temporal dependency

Fig. 2

a Center views of AICS cells labeled with microtubules, imaged using sLFM at low SNR in simulation, before and after LF-denoising. The image bit depth is set to 16, Gaussian noise variance to 5, and the photon number of maximum intensity to 100. b Enlarged regions obtained by different denoising methods, including BM3D, Noise2Void, Noise2Noise, DeepInterpolation, DeepCAD-RT, DeepSeMi, SRDTrans and LF-denoising. Two example angular views are shown. High-SNR data is regarded as ground truth. Yellow and blue arrows indicate the image blur and detail loss. c Another enlarged regions after reconstruction, with insets showing corresponding Fourier transforms. The normalized intensity profiles and modulation transfer functions (MTFs) of white arrows in the left most image across adjacent tiny structure are shown for comparison. The presence of a valley in the middle of the MTF curve indicates strong resolving capability of LF-denoising for spatially adjacent structures. d Boxplots showing comparisons between different denoising methods using RMSE (lower is better), SSIM, PNSR and Pearson correlation (higher is better). Metrics are calculated on data after reconstruction. The boxplot formats: center line, median; box limits, lower and upper quartiles; whiskers, 1.5-fold interquartile range. All P values were calculated using one-sided independent t test, significance at P < 0.05. *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001. n = 30 denotes the number of individual samples. P values are 1.71 × 10−19, 9.03 × 10−19, 7.35 × 10−25, 3.55 × 10−23, 2.71 × 10−15, 3.51 × 10−24 and 1.28 × 10−11 from left to right for RMSE, and 2.95 × 10−19, 9.84 × 10−19, 6.98 × 10−33, 2.48 × 10−17, 4.72 × 10−23, 2.38 × 10−28 and 2.49 × 10−12 from left to right for SSIM, and 9.02 × 10−25, 1.35 × 10−22, 6.85 × 10−39, 3.02 × 10−27, 8.39 × 10−16, 3.29 × 10−35 and 3.43 × 10−17 from left to right for PSNR, and 3.76 × 10−18, 2.51 × 10−18, 1.89 × 10−22, 1.63 × 10−23, 8.43 × 10−18, 5.41 × 10−25 and 1.33 × 10−10 from left to right for Pearson correlation. The asterisks indicate the significant levels of comparisons between each method and LF-denoising. Scale bars, 5 μm (a–c) and 2 μm−1(c).

Back to article page