Table 1 Comparison of micrograph denoising methods based on estimated SNR (in dB, larger is better).

From: Topaz-Denoise: general deep denoising models for cryoEM and cryoET

Method

EMPIAR-1026148

EMPIAR-1000524

EMPIAR-1002549

Protocadherin (K2)

18sep08d (K2)

19jan04d (K2)

19may10e (K2)

18aug17l (Falcon III)

18sep06d (Falcon III)

18sep19l (Falcon III)

Overall

Affine (Topaz)

5.49

1.29

0.72

4.83

4.51

8.87

12.02

10.65

6.90

9.15

6.44

U-net (Topaz)

7.17

1.72

1.07

5.94

6.06

8.43

13.07

15.17

7.37

13.24

7.92

Low-pass

5.19

−0.12

−0.40

4.22

3.53

6.87

9.99

9.04

6.95

8.71

5.40

Raw

−17.14

−20.13

−24.15

−14.47

−15.40

−11.73

−5.44

−6.33

−3.64

−5.63

−12.41

  1. SNR was estimated from 20 paired signal and background regions selected for each dataset. In each column, the best performing model is bolded. We report denoising results on aligned micrographs for the NYSBC K2 and Falcon III datasets. All datasets were collected in electron counting modes, except for 18sep06d, which was collected using Falcon III integrating mode. Our U-net denoising model performs best overall and is best on all except for the 19jan04d dataset where our affine denoising model slightly outperforms it. We report low-pass filtering by a binning factor of 16 on all datasets, which we found to give better SNR overall compared to Gaussian low-pass filtering.
Back to article page