Denoising

Improving image quality and SNR with denoising

Bioimages are often corrupted by noise, which reduces image quality and the signal-to-noise ratio (SNR). Especially when using low laser light intensities in fluorescence microscopy (e.g. to be gentle to live specimen), noise can become very large and render data interpretation difficult. As this is a very general problem, several deep learning approaches were developed to denoise microscope images. This includes supervised networks and self-supervised networks. In DeepBacs we used CARE and Noise2Void as supervised and self-supervised networks, respectively.

Important disclaimer

Here are the links to the original articles/repositories of the networks we employ for denoising:

CARE:

• Content-aware image restoration: pushing the limits of fluorescence microscopy by Martin Weigert et al.
• CARE original code and documentation is freely available in GitHub

Noise2Void (N2V):

• Learning Denoising from Single Noisy Images by Alexander Krull, Tim-Oliver Buchholz and Florian Jug
• Noise2Void original code and documentation are freely available in GitHub.

Please also cite these original papers when using CARE or Noise2Void with our notebooks.

Datasets

Denoising of confocal fluorescence images of live E. coli

Data generation

E. coli strain CS01 carrying a chromosomal H-NS-mScarlet-I protein fusion (parental strain NO34, kind gift from Zemer Gitai) was grown in LB Lennox at 25°C and shaking at 220 rpm. To generate the training dataset, cells were fixed chemically using a mixture of 2% formaldehyde and 0.1% glutaraldehyde (as described in Spahn et al, 2018). Fixed or live cells were immobilized under agarose pads poured into gene frames following the protocol by de Jong et al.. Imaging was performed on a commercial Leica SP8 confocal microscopy (Leica Microsystems) bearing a 1.40 NA 63x oil immersion objective (Leica Microsystems). To increase optical sectioning, the pinhole size was set to 0.5 AU and 512 x 512 px² confocal images (45 nm pixel size) were recorded and emission was detected with HyD detectors in standard operation mode (gain 100, detection window 570 – 650 nm). For the training dataset, a two-channel image of the same structure was recorded in frame sequential mode using different settings for low (0.03% 561 nm laser light, no averaging) and high SNR images (0.1% 561 nm laser light, 4x line averaging), respectively. For live-cell time series, the field of view was reduced to 256 x 256 px² to allow for fast acquisition of high SNR images at ~ 0.8 Hz. Low SNR time series were recorded at similar frame rate by including a lag time.

Example images of the paired training data used to train the CARE model:

Network training parameters

Networks for denoising of live-cell nucleoid dynamics in E. coli were trained using the following parameters:

Network	Images (train/test)	Epochs	Steps	Image size	Patch size (#/img)	Batch size	LR	% Valid.	Augmentation	Train time
CARE	28/2	100	600	512 x 512 px²	64 x 64 px² (50)	8	0.0004	10	4x (flip/rot.)	30 min 34 s
N2V	28/2	200	101	512 x 512 px²	64 x 64 px² (50)	128	0.0004	10	4x (flip/rot.)	41 min

Denoising of VerCINI images of FtsZ in live B. subtilis

Data generation:

The raw data analysed here were acquired and analysed as presented in Whitley et al. 2021. Cells were trapped in microhole-containing agarose pads (6%) that were poured on nanofabricated silicon micropillars (see paper linked above). Silicone micropillar wafers were nanofabricated and used to prepare agarose microholes as described previously (Whitley et al., 2021). The agarose pad was transferred into a gene frame, and agarose surrounding the micro-hole array was cut away. Concentrated liquid cell culture at mid exponential phase (OD600 ~ 0.4) was loaded onto the pad and centrifugation using an Eppendorf 5810 centrifuge with MTP/Flex buckets loaded individual cells into the microholes. The pad was then washed to remove unloaded cells. This repeated several times until a sufficient level cell loading was achieved. Cells were imaged at 1 frame/second with continuous exposure for 2 minutes at 1-8 W/cm2. Lateral drift was then corrected using StackReg.

Network training:

As Noise2Void can be used on the image to be denoised itself, we used the first image frame of the time series we aimed to denoise.

Training was performed using the following parameters:

Network	Images (train/test)	Epochs	Steps	Image size	Patch size (#/img)	Batch size	LR	% Valid.	Augmentation	Train time
N2V	3/n.a.	200	44	1024 x 1024 px²	64 x 64 px² (500)	128	0.0004	10	4x (flip/rot.)	18 min

Training in google colab

Network	Link to datasets and pretrained models	Direct link to notebook in Colab
CARE (2D)	E. coli nucleoid dynamics
N2V (2D)	B. subtilis FtsZ dynamics

LINKS STILL HAVE TO BE ADDED