We demonstrate the use of a simulation supervised machine learning tool for analyzing mitochondria in fluorescent microscopy images of fixed cells. Current methods to segment mitochondria in microscopy images include automatic thresholding based methods such as Ostu or manual segmentation. Thresholding based techniques perform poorly where the signal to background ratio is low.
Typically, images for the morphological analysis feature a large number of mitochondria, making manual segmentation tedious. Supervised deep learning methods perform segmentation with high accuracy, but require a large number of input ground-truth pair data for training. This approach uses a physics based simulator to generate pairs of microscopy images, and their 2D ground-truth shape mask, thus eliminating the need for manual annotation.
The model trained on simulated images is then used to segment mitochondria in real microscope images. We use a deep learning based segmentation tool that enables automation of this task with high accuracy and that does not require an annotated ground-truth dataset for training. The simulation starts with the geometry generation using parametric curves for shape generation.
Emitters are uniformly distributed and randomly placed on the surface of the generated shape such that the density matches experimental values. A 3D PSF of the microscope is computed using the Gibson-Lanni model. To achieve photorealism between the simulated images and the experimental images, we emulate both the dark and shot noises.
The physics ground-truth is generated in the form of a binary map. We assess the effect of CCCP treatment on mitochondria morphology of fixed cardiommyoblasts imaged using a confocal microscope. Cell culture.
Prepare for seeding the cells, operating in a sterile laminar flow hood, by placing a cover slip for each experimental condition in a well of a 12 well plate. Ensure the diluted cell suspension is properly mixed by pipetting the contents of the centrifuge tube up and down several times before dispensing the appropriate volume to each well. Experimental procedure.
Aspirate cell culture media from the wells of the 12 well plate then quickly apply fresh preheated media to the control wells and the preheated media with 10 micromolar CCCP solution to the test condition wells. Incubate in the 37 Celsius cell incubator for two hours. Aspirate cell culture media from the wells and apply the preheated fixation solution.
Staining and mounting of cells on cover slips. Add 10 microliter mounting media to a prepared glass slide to mount a cover slip. Pick up the cover slip from the 12 well plate using tweezers and dab moisture off the cover slips by briefly touching the edge and the back of the cover slip to the prepared lint free paper towel.
Gently lower the cover slip down on the waiting droplet of mounting media. Microscope and imaging. Using the oculars, manually adjust the Z level to have the sample in focus.
Switch to the acquired tab within software. Use the smart setup to select fluorescence channel to be used for imaging. Array-centered on the middle of the cover slip with 12 total positions to be imaged.
Position RA centered on the center of the cover slip. It's possible to image multiple locations in an automated fashion. Generating simulated training data.
Download the code and unzip the contents. Follow the instructions and readme to set up the environment. Navigate to the folder named src"Make a copy or use the folder 2.
mitochondria simulation airy"and rename it. This folder contains all the files related to the simulation of the training data. There are three sets of parameters to be set for the simulation.
First, set the parameters of number of mitochondria, range of diameter, range of Z-axis, and density of fluorophores for the simulator in the batch config file. Next, set optical parameters of numerical aperture, magnification, and minimum wavelength of the dataset in the file microscopePSFmode. py"Set the desired value for pixel size and emission wavelength in the fire generate_batch_parallel.
py"Set the third set of parameters regarding the output dataset, like the size of output images, the number of tiles in each image, and number of total images in the file generate_batch_parallel. py"Run the file generate_batch_parallel. py"to start the simulation.
To obtain the final sized image, make a copy of the folder named 5. Data preparation and training/data preparation"and navigate into it. Set the parameters of batch number and the number of images per batch, range of noise to be added in the file Data_generator.
py"Run the file Data_generator. py"create the montage images. Copy the folders named image"and segment"into the datatrain/train"folder.
Deep learning based segmentation. For training the segmentation model for a new microscope image setting, navigate to the folder named train"and set the parameters of bach size, backbone model for the segmentation, the number of epochs, and learning rate for training inside the file named train_unet. py"Run the file train_unet.
py"to start the training. The training process displays the metric for evaluating the segmentation model on the simulated validation set. After the training is complete, the model is saved as best_model.
h5"in the folder named train"To test the model on the microscope images, the images must be split to the size desirable by the train model. For this, navigate to the folder named 6. Prepare test data"and copy the PNG format files of the data into the folder PNG"and run the file split_1024_256.
py"This will create 256 by 256 size crops of the images in the data folder. To segment image crops, go to the folder named 7. Test segmentation"and run the file named segment.
py"after setting the name of the saved model to be used. The segmented images are saved into the output folder. Morphological analysis.
Place the file named make_montage. py"into the folder named 7. Test segmentation"and run the file to stitch the segmented output back to the original size of the image.
Create a new folder named 9. Morphological Analysis"in the source folder and navigate into it. Install the skan and seaborn libraries using the command pip install seaborn skan"The segmentation masks are skeletonized using the library named skan"to enable analyzing the topology of the individual mitochondrion.
Place the file into the folder 9. Morphological Analysis"Arrange the images of different groups of the experiment into different folders inside the folder 7. Test segmentation"Run the file to create plots for the analysis.Results.
The quantitative analysis to be performed depends on the research questions or hypothesis. In our experiment, we were interested in the three different morphologies of mitochondria, namely dots"small mitochondria bits, rods"fiber-like or string like mitochondria, and multi branched networks"To identify the morphology, we first skeletonize the segmentation output and analyze the branch links of the different classes. We show the results of the analysis for the two groups of galactoses adapted cells, the control group and the CCCP treated group.
We see a significant increase in the mean branch lengths of dots, which is expected, given the swelling mitochondria undergo when exposed to CCCP. The percentage of mitochondria in individual lengths of both rods and network classes is significantly reduced relative to the control when the cells have been treated with CCCP, thus verifying our hypothesis. A challenging scenario for our method is when the image has densely packed mitochondria.
The pink color indicates the longest single mitochondrion detected in the image, which is caused due to the increased error in the segmentation results. These failure cases can be detected by a control of the lengths of mitochondria and using morphological operators. While our application is an example of how we performed analysis of mitochondrial morphology, we believe that such an analysis and research questions around it can be formulated for various aspects of mitophagy and related biological experiments.
