�Segmentation and Statistical Analysis of Cellular Images using Deep-Learning��Prof. Abdul Barakat

1

Reference: Pachitariu, M., Stringer, C. Cellpose 2.0: how to train your own model. Nat Methods 19, 1634–1641 (2022). 

Stringer, C., Wang, T., Michaelos, M. et al. Cellpose: a generalist algorithm for cellular segmentation. Nat Methods 18, 100–106 (2021).

Thenier, F. Segmentation of Microfluidic Images using Deep-Learning - Focusing on Image enhancement and domain transfer challenges�

0. Introduction

2

4

https://www.jove.com/fr/v/10511/growth-curves-generating-growth-curves-using-colony-forming-units?language=Chinese

Dataset

1. Preprocess images

5

6

Preprocess

Limit noise amplification

Deepen the gap between background and nuclei intensity values

CLAHE

Sigmoid

Reference: Thenier, F. Segmentation of Microfluidic Images using Deep-Learning - Focusing on Image enhancement and domain transfer challenges�

Preprocessing

Image quality metrics

SSIM (Structural Similarity) :

measure the visual quality of a compressed or denoised image compared to the original image, by measuring the structural similarity between the two images.

The closer to 1, the better.

PSNR (Peak Signal-to-Noise Ratio) :

PSNR is a metric used to measure the quality of reconstructed images compared to the original,

where a higher value indicates better resemblance.

Note: neither metric is perfect, and the best one to use depends on the specific application and characteristics.

9

CLAHE parameters

Sigmoid parameters

2. Train the model

10

After that, we randomly took 25 images from each cell, completed the annotation.

MCF10A

0

1

2

3

4

MCF10A_Braf

MCF10A_Rac

Annotation

0

1

2

3

4

0

1

2

3

4

12

After that, I trained the model with different numbers of images, with or without preprocessing and tested their accuracy.

Pachitariu, M. Nat Methods 2022

Conclusion:

Preprocessing does not seem to improve the accuracy of the model.

​

It is also in line with the paper: only 5 images are needed to train a more accurate model.

3. Use the model to segment

17

Use custom model to segment images

​

Sources: https://forum.image.sc/t/cellpose-flow-and-cell-threshold/70347/4

Flow Threshold:

It determines the maximum allowed error of the predicted flows for each mask. It's used to ensure the recovered shapes after the flow dynamics step are consistent with real ROIs.

​

Cell Probability Threshold:

It determines the minimum cell "probability" required for pixels to be used in dynamics determination and ROI identification.

Experiments about

Threshold: Expen Phase

B1_3_frame_0022

In the case of using the trained model,

​

• Flow threshold 0.4 and Cell prob 0 yield high average precision.

Remark

4. Analysis the results

23

We ran Cellpose on each of the well images, counted the number and area change with time per well for each cell.

Statistic

MCF10A

0

1

2

3

4

MCF10A_Braf

MCF10A_Rac

0

1

2

3

4

0

1

2

3

4

MCF10A_0

MCF10A_2

MCF10A_3

MCF10A_4

MCF10A_1

MCF10A_Braf_0

MCF10A_Braf_2

MCF10A_Braf_3

MCF10A_Braf_4

MCF10A_Braf_1

MCF10A_Rac_0

MCF10A_ Rac_2

MCF10A_ Rac_3

MCF10A_ Rac_4

MCF10A_ Rac_1

1. Cell area both increase and then decrease with time. It can be fitted with a cubic function. And cell numbers can be fitted with a logistic function.

​

​

​

• MCF10A and MCF10A_Braf have similar cell area distribution and number changes. While MCF10A_Rac has a larger cell area and a more dispersed cell area over time, but also has fewer cells.

​

​

Remark:

5. Convolutional Autoencoder

33

Unsupervised learning classification

More work to optimize the network structure and classifier performance needs to be done in the future…

Thank you very much!

36