Fig 1.
Comparison of noise reduction and blur.
A: Noisy TEM image of cilia. B: Comparison of an image section of the noisy image, and two denoised versions. While method A reconstructs a sharper image compared to method B (blur effectA = 0.34, blur effectB = 0.68, lower is better), method B reduces noise more strongly (PSNRA = 28.29, PSNRB = 32.44; higher is better). C: Segmentation masks generated with Otsu’s method [11] based on the noisy image and the reconstructions of method A and B.
Fig 2.
Results for blind zero-shot denoising TEM images of SARS-CoV-2 infected cell cultures.
A: Qualitative comparison of denoised image sections showing virus and cell structures obtained with the investigated algorithms (image sections are scaled locally to fill the color range). B: Quantification of the denoising performance in terms of SNR in dB based on hand-labeled signal and background regions and blur effect (SNR and blur effect plotted against each other). The left subfigure displays the performance of all algorithms, while the right subfigure provides an enlarged view of the region populated by most algorithms. For both the SNR and blur effect measure, we measured the difference (delta) between the values obtained for the reconstructed and the noisy image and here depict averages together with standard errors of the mean (SEM) over the nine considered test images.
Fig 3.
Results for blind zero-shot denoising short exposure TEM images of cilia (benchmark adopted from Bajic et al. [42]).
A: Qualitative comparison of denoised image sections showing a cilium (image sections chosen to match Fig 3 in [42]; image sections are scaled locally to fill the color range). B: Quantification of the denoising performance in terms of PSNR in dB and the BAD of the pseudo ground-truth (PGT) image and the denoised reconstruction (PSNR and BAD values plotted against each other). For the PSNR measure, we computed the difference (delta) between the values obtained for the reconstructed and the noisy image and here depict averages together with standard errors of the mean (SEM) over three independent executions of the experiment (note that the SEM values are very small and barely visible in the plot and that they were only evaluated for stochastic algorithms). Points are marked in red or blue if the blur effect value of the PGT image is larger or smaller than the blur effect value of the reconstructed image, respectively. Further details are discussed in the text.
Fig 4.
Results for blind zero-shot denoising of fluorescence microscopy images (benchmark adopted from Prakash et al. [30]).
A: Qualitative comparison of denoised image sections (image sections and colormap chosen to match Fig 2 in [30]; image sections are scaled locally to fill the color range). B: Quantification of the denoising performance in terms of PSNR in dB and the BAD of the pseudo ground-truth (PGT) image and the denoised reconstruction (PSNR and BAD values plotted against each other). For the PSNR measure, we computed the difference (delta) between the values obtained for the reconstructed and the noisy image and here depict averages together with standard errors of the mean (SEM) over three independent executions of the experiment (note that SEM were only evaluated for stochastic algorithms). Points are marked in red or blue if the blur effect value of the PGT image is larger or smaller than the blur effect value of the reconstructed image, respectively. Further details are discussed in the text.
Table 1.
The noise reduction algorithms considered have different strategies for noise removal and types of noise assumptions.
Median filtering is excluded from this table since it has no explicit assumption regarding the noise being additive or signal-dependent. While the DivN approach, in principle, allows for incorporating any suitable imaging noise distribution, here we have follow the publicly available implementation and used a variant of the algorithm in which the noise distribution is described by a GMM. BM3D and ES3C incorporate a Gaussian noise assumption; the VST-based approaches and PMM assume Poisson noise.