1 of 32

Noise2Stack: �Improving Image Restoration by Learning from Volumetric Data

Mikhail Papkov, Kenny Roberts, Lee Ann Madissoon, Jarrod Shilts, Omer Bayraktar, Dmytro Fishman, Kaupo Palo, and Leopold Parts

MLMIR 2021

1

01.10.2021

2 of 32

Agenda

  • Motivation for the image restoration task
  • Dataset description: useful properties of volumetric data
  • Denoising with neural networks
  • Noise2Stack: what is different?
  • Results and discussion
  • Conclusion

2

3 of 32

Motivation

3

4 of 32

No data is absolutely clean

...especially in biology:

    • Acquisition speed trade-off
    • Imaging depth
    • Phototoxicity

4

5 of 32

Noise is problematic

...for both manual annotation and existing automated pipelines (such as pre-trained neural networks), because

5

Garbage in

Garbage out

But we can do better

Segmentation

6 of 32

Data

6

7 of 32

MRI dataset

  • IXI (Information eXtraction from Images) T1 dataset
  • Clean 256x256 px images from 60 patients, 150 planes in stack (100 middle used)
    • 48/2/10 train/val/test split
    • Scaling to [-0.5, 0.5]
  • Used in previous work on denoising

7

8 of 32

Noise generation

 

8

FFT

Sample

10%

IFFT

9 of 32

Post-processing opportunity

All the frequencies that are present in the image are there for sure (because of the frequency sampling procedure) — copy-paste them to the output spectrum

  • Not applicable to other types of noise!

9

Ctrl-C Ctrl-V

Input spectrum

Denoised spectrum

10 of 32

Microscopy

  • In-house dataset: 270 images 2160x2160 px, 10 focal planes in z-stack
    • 180/30/60 train/val/test split
    • Percentile normalization

  • Two modalities
    • Fluorescence: low exposure (20 ms) and high exposure (100 ms) registered pairs
    • Brightfield: only one exposure

10

11 of 32

Denoising with neural networks

11

12 of 32

U-Net

An encoder-decoder architecture with skip connections, widely used for denoising and segmentation tasks (hyperparameters could vary)

Different configurations for MRI and microscopy from the literature

12

13 of 32

U-Net

The exact architecture is heavier

  • 5 downsampling steps
  • 1 conv per step down
  • 2 conv per step up
  • Concatenate input
  • Heavy decoder

13

14 of 32

Noise2Clean

Use clean (noise-free, high-resolution) images as training targets (ground truth).

Problem: clean ground truth is a rare thing in biomedical imaging.

14

Input

Output

Target

MSE

15 of 32

Noise2Noise

Use independent noisy acquisitions of the same image as both input and target

Problem: many existing datasets do not have a second copy

15

Input

Output

Target

MSE

16 of 32

Noise2Noise

For MSE loss the minimum is found in expected value of each pixel

  • zero-mean noise cancels out

16

Input

Output 1

Output 1

Output N

Target

MSE

Training

17 of 32

Noise2Noise as a baseline

Previous works show that Noise2Noise performs on par with traditional denoising (Noise2Clean) and outperforms existing self-supervised methods (e.g., Noise2Void)

Self-supervised methods does not work with structured noise

17

18 of 32

Noise2Stack

18

19 of 32

Noise2Stack

In Noise2Noise every plane in a stack is treated independently on its position

19

20 of 32

Noise2Stack

In Noise2Stack we sample neighbors along with the target plane (copy-supervised mode) or without it (self-supervised mode)

20

21 of 32

Results

21

22 of 32

Metrics

  • PSNR = peak signal-to-noise ratio

  • SSIM = structural similarity index

  • NRMSE = normalized root mean squared error

22

23 of 32

MRI

23

24 of 32

MRI

24

Input

Noise2Noise baseline

Noise2Stack (self-supervised)

Noise2Stack (copy-supervised)

Ground truth

Stack

25 of 32

Microscopy

25

26 of 32

Downstream segmentation

26

27 of 32

Plane denoising

In MRI, middle planes are harder to denoise (less background, symmetry point).

In microscopy, marginal planes are noisier.

27

28 of 32

Conclusion

28

29 of 32

Conclusion

  • Neighboring planes in dense volumetric stacks contain a lot of shared information that can be used for denoising
  • Noise2Stack achieves the state-of-the-art denoising performance in the copy-supervised mode
  • Noise2Stack in the self-supervised mode performs on par with Noise2Noise providing an opportunity to achieve high quality results with a single stack acquisition
  • The presented method can potentially be applied for any kind of volumetric images, the exact data properties and other limitations are the subject of further studies.

29

30 of 32

Acknowledgements

This work was funded by PerkinElmer Inc. (VLTAT19682) and Wellcome Trust (206194). We thank High Performance Computing Center of the Institute of Computer Science at the University of Tartu for the provided computing power.

30

31 of 32

31

32 of 32

Plane similarity in volumetric images

Neighboring planes have a lot in common

Can this shared information be used for denoising?

32