This code repository accompanies the article titled "A Generative Benchmark for Evaluating the Performance of Fluorescent Cell Image Segmentation." The toolkit, outlined in this repository, serves two main purposes: it facilitates the generation of diverse cell contours using a StyleGAN2-based approach and enables realistic contour rendering through the utilization of Pix2PixHD.
- Cell Contour Generation: The toolkit employs StyleGAN2 to generate diverse and graded-density cell contours.
- Contour Rendering: Leveraging Pix2PixHD, the repository enables realistic rendering of cell contours.
- Linux and Windows are supported, but we recommend Linux for performance and compatibility reasons.
- 1–8 high-end NVIDIA GPUs with at least 12 GB of memory.
- GCC 7 or later (Linux) or Visual Studio (Windows) compilers. Recommended GCC version depends on CUDA version, see for example CUDA 11.6 system requirements.
- 64-bit Python 3.8 and PyTorch 1.8.2 (or later). See https://pytorch.org for PyTorch install instructions. CUDA toolkit 11.0 or later.
- Python libraries: Install the required libraries by running
pip install -r requirements.txt.
- Clone this repo:
git clone https://github.com/edwardcao3026/SegBenchmark.git
cd SegBenchmark-
Trained networks are stored as
*.pklfiles that can be referenced using local filenames. -
Generate diverse cell contours using the following command:
python generate_contours.py --outdir=./output_folder --seeds=number of seeds --network= path to netwrok - Utilize
image_rendering.pyfor realistic image rendering:
python image_rendering.py --label_nc 0 --no_instanceAt this stage, we will use the generated dataset to compare the performance of various segmentation algorithms. This process involves inputting the generated dataset into different segmentation algorithms and evaluating their performance and effectiveness. By comparing different algorithms, we can find the segmentation algorithm that is most suitable for a specific task, which can help us improve the accuracy and efficiency of segmentation in practical applications.
Here are the specific steps for using three different segmentation methods in articles:
# Install cellpose and the GUI dependencies from your base environment using the command
python -m pip install cellpose[gui]
# The quickest way to start is to open the GUI from a command line terminal.
python -m cellpose- Load an image in the GUI (by dragging and dropping the image or by selecting "Load" from the File menu).
- Set the model: CellPose has two models, cytoplasm and nuclei, which correspond to the segmentation of cell cytoplasm and cell nuclei, respectively.
- Set the channels: Select the image channel to be segmented. If segmenting cytoplasm, select the green channel; if segmenting nuclei, select the red/blue channel. If there are both cytoplasm and nuclei in the image, set chan to the channel where cytoplasm appears, and set chan2 to the channel where nuclei appear. If segmenting cytoplasm but there is no nucleus, only set chan and leave chan2 as None.
- Click on the calibrate button to estimate the size of objects in the image. Alternatively, manually input the cell diameter for calibration. The estimated size will be reflected by a red circle in the lower left corner.
- Click on the "run segmentation" button to start the segmentation process. You can choose to display the segmentation mask by checking the "MASKS ON" option.
The parameter inputs in the GUI interface can also be achieved through the terminal mode:
python -m cellpose --dir ~/images_cyto/test/ --pretrained_model cyto --chan 2 --chan2 3 --save_pngAll parameters can be viewed using the help parameter:
python -m cellpose -hSimilar to the previous two methods, CellPose can also be called directly in Python code for programming:
from cellpose import models
import skimage.io
model = models.Cellpose(gpu=False, model_type='cyto')
files = ['img0.tif', 'img1.tif']
imgs = [skimage.io.imread(f) for f in files]
masks, flows, styles, diams = model.eval(imgs, diameter=None, channels=[0,0],
threshold=0.4, do_3D=False)3.2 CellPofiler [2]
CellProfiler is a free software developed by the Broad Institute of Harvard and MIT. It is designed to enable biologists to quantitatively measure phenotypes of thousands of images automatically, without the need for computer vision or programming training.
Creating a specific pipeline for cell segmentation in CellProfiler involves a series of steps and modules tailored to this task. Here's a outline of what such a pipeline might look like:
-
Load Images:
- Use the
Imagesmodule to specify the images you want to analyze.
- Use the
-
Identify Primary Objects:
- Employ the
IdentifyPrimaryObjectsmodule. This is crucial for cell segmentation, as it identifies individual cells in your images. Common settings involve specifying the typical size of the cells and the intensity threshold for segmentation.
- Employ the
-
Identify Secondary Objects:
- If your analysis requires identifying objects surrounding or attached to the primary objects (like cytoplasm around a nucleus), use
IdentifySecondaryObjects. This module relies on the primary objects as a reference.
- If your analysis requires identifying objects surrounding or attached to the primary objects (like cytoplasm around a nucleus), use
-
Measure Object Size and Shape:
- The
MeasureObjectSizeShapemodule measures various properties of the identified objects, such as area, perimeter, and form factor.
- The
-
Export Data:
- Finally, use
ExportToSpreadsheetandSaveImagesmodules to export your results for further analysis.
- Finally, use
3.3 DeepCell [3]
Researchers have developed a deep learning segmentation algorithm, Mesmer, which consists of a ResNet50 backbone and a feature pyramid network. It automatically extracts key cell features, such as subcellular localization of protein signals, achieving human-level performance.
Visit the pre-trained deep learning models on DeepCell.org. This website allows you to easily upload example images, run them on available models, and download the results without any local installation required.
-
Install DeepCell with pip:
pip install deepcell
-
Using DeepCell with Docker:
- If you have a GPU, ensure you have CUDA and Docker installed.
- Run the Docker command to start a container with DeepCell installed:
docker run --gpus '"device=0"' -it --rm -p 8888:8888 -v $PWD/notebooks:/notebooks -v $PWD/data:/data vanvalenlab/DeepCell-tf:latest-gpu
- This command starts a Jupyter session and mounts data and notebook directories.
-
Example Usage:
- The DeepCell documentation includes examples of training segmentation and tracking models.
- You can find Python notebooks for these examples, illustrating how to use DeepCell for single-cell analysis.
For more detailed information and examples, you should refer to the DeepCell documentation. This resource provides comprehensive guidance on installing and using DeepCell, including example notebooks for various applications.
To run DeepCell in Python, you can follow this script:
import numpy as np
from skimage.io import imread
from deepcell.applications import Mesmer
app = Mesmer()
# List of image file paths
files = ['image1.jpg', 'image2.jpg']
for file in files:
im = imread(file)
# Select only the relevant channels (assuming the image has multiple channels)
im = im[:, :, [1, 2]]
# Expand dimensions to match the input shape expected by Mesmer
im = np.expand_dims(im, axis=0)
# Predict the labeled image
labeled_image = app.predict(im, compartment='both')After running the three segmentation methods as described above, you can proceed to analyze the segmentation results by running the analysis.py script. The results of the analysis will be saved as a CSV (Comma-Separated Values) file.