The use of Artificial Intelligence (AI) for analysis of cellular images is very promising, especially for image segmentation tasks that were previously challenging or unfeasible, such as the segmentation of nuclei on brightfield images.
Brightfield images, while being widely accessible, often suffer from low contrast and lack the specific staining that fluorescent images provide.
This application note showcases how an AI-based model has been trained to segment brightfield images and how the pre-trained AI model was directly applicable to unknown cell lines.
Download this application note to discover:
- A ready-to-use pre-trained AI model compatible with different magnifications and cell lines
- How this AI model provides robust segmentation of cells on brightfield images
- Digital phase image reconstruction for cytoplasm segmentation
Phenologic.AI:
Nuclei segmentation
on brightfield
images using a
pre-trained
Artificial Intelligence
(AI) model.
Introduction
The use of Artificial Intelligence (AI) for analysis of cellular images
is very promising, especially for image segmentation tasks that
were previously challenging or unfeasible. The segmentation of
nuclei on brightfield images is a prime example of such a task.
Brightfield images, while being widely accessible, often suffer
from low contrast and lack the specific staining that fluorescent
images provide.
To create an effective AI-based module for cell segmentation,
the neural network needs to be trained with many ground truth
images that need to be representative of the actual sample
images. However, the segmentation quality may drop if a
different cell type or a different magnification has been used
and the images look different from experiment to experiment.
Many current AI-based methods therefore require users to
re-train their models, sometimes even by manual drawing on
images, e.g., to mark nuclei or cells, which is time-consuming
and cumbersome. Key points
• Brightfield images are rich in information
but difficult to segment
• Phenologic.AI™ provides robust
segmentation of cells on brightfield images
• Ready-to-use pre-trained AI model
compatible with different magnifications
and cell lines
• Phenologic.AI performs digital phase image
reconstruction for cytoplasm segmentation
TECHNICAL NOTE
Phenologic.AI: Nuclei segmentation on brightfield images using a pre-trained Artificial Intelligence (AI) model.
www.revvity.com 2
Here we show results for 16 different cell lines and for two
commonly used magnifications (10x and 20x).
Of the 16 cell lines only two - A549 and MCF7 - have been
part of the training data set for the AI model, while 14
are “unknown” to the model. Furthermore, the data used
here for A549 and MCF7 are from different experiments
(cell seeding at a different lab using a different stock) than
the training data.
Phenologic.AI also allows to reconstruct digital phase
images if two planes are acquired. Examples of two cell
lines with different morphologies are shown in Figure 2.
Pre-trained deep-learning
image-analysis
Phenologic.AI has been trained on a diverse dataset of
thousands of images from various cell lines, captured with
different objectives. This extensive training has endowed
it with a high level of universality, overcoming the need
for training by the scientist. Additionally, automated plane
selection and dedicated AI-building blocks enhance the
ease-of-use of Phenologic.AI (Figure 1).
Figure 1: The automated selection of two brightfield planes, along with dedicated building blocks for nuclei segmentation and digital phase
image construction for cytoplasm detection, enhances the ease-of-use of Phenologic.AI.
Phenologic.AI: Nuclei segmentation on brightfield images using a pre-trained Artificial Intelligence (AI) model.
www.revvity.com 3
In addition to brightfield, PhenoVue™ Hoechst 33342 images
were acquired and analyzed using the “gold standard”
Find Nuclei building block of Harmony™ software as a
reference. This segmentation was used as the ground truth
and false positive detection by Phenologic.AI identified by
the lack of overlap between both segmentations. Figure 4
shows the results of the detection rate (ratio of Phenologic.
AI detected nuclei / ground truth nuclei) and the percentage
of false positive nuclei.
For most cell lines and objectives, detection rates above
0.9 were achieved, with less than 4% false positives.
Figure 3: Phenologic.AI Building Block – Find Nuclei (AI) The model
allows to switch between 1 plane and 2 plane model for nuclei
detection. Upper and lower plane need to be acquired in separate
channels (not as stack).
The new Phenologic.AI building block is available in
Signals Image Artist™ and is shown in Figure 3.
Figure 2: Example images of segmentation results of two cell lines: A549 cells being equally distributed with well separated nuclei,
and BxPC3 cells as an example for insular growth pattern and narrower cells. The first row shows the ground truth nuclei segmentation
on Hoechst channel, and the second row the AI-based nuclei segmentation based on brightfield images. The last row is an example for a
cytoplasm segmentation using an AI-based digital phase contrast image.
Phenologic.AI: Nuclei segmentation on brightfield images using a pre-trained Artificial Intelligence (AI) model.
www.revvity.com 4
Figure 4: The detection rate of the brightfield AI model on 20xWI (water immersion), 20xLWD (long working distance) and 10x objective images
was above 0.9 for all cell lines. (A) The ratio of the number of detected nuclei in brightfield to those in fluorescent channel yielded values
above 0.9 for both 10x and 20x objectives across most cell lines. (B) For most cell lines, the percentage of false positive nuclei was below
4%. Error bars indicate standard deviation, n ≥ 93 wells.
Cell Type, Objective
Detection rate [-] False positive rate [%]
Cell Type, Objective
Phenologic.AI: Nuclei segmentation on brightfield images using a pre-trained Artificial Intelligence (AI) model.
www.revvity.com
Copyright ©2024, Revvity, Inc. All rights reserved. For research use only. Not for use in diagnostic procedures. 1565257
Conclusion
Here, we have shown that the pre-trained AI model was
directly applicable to unknown cell lines. For most of
the cell lines and objectives, the model yielded ratios of
AI-detected nuclei to ground truth nuclei well above 0.9
and false positive rates below 4%. This underscores the
universality of the model’s usage, alleviating the time and
computational demands of model training. This opens up
new application workflows for both live and fixed cell
applications. Cells can easily be segmented into nuclei
and cytoplasm (in combination with the AI-generated DPC
channel), allowing not only the use of the cell area as a
surrogate marker for cell proliferation but also the counting
of individual cells and relating this to cell morphology
parameters if needed. Since no fluorescent dye is needed
for nuclei or cytoplasm detection, this approach is
extremely gentle to cells, allowing short imaging intervals
with minimal disturbance. For fixed cell applications, this
allows the omission of widely used Hoechst or DAPI staining,
enabling the use of another target-specific stain, such as
PhenoVue Fluor 405 – Phalloidin.
