Observing individual cells through microscopes can reveal a variety of vital cell biological phenomena that incessantly play a task in human diseases, however the means of distinguishing single cells from one another and their background is incredibly time consuming — and a task that’s well-suited for AI assistance.
AI models learn methods to perform such tasks by utilizing a set of information which are annotated by humans, however the means of distinguishing cells from their background, called “single-cell segmentation,” is each time-consuming and laborious. In consequence, there are limited amount of annotated data to make use of in AI training sets. UC Santa Cruz researchers have developed a technique to unravel this by constructing a microscopy image generation AI model to create realistic images of single cells, that are then used as “synthetic data” to coach an AI model to raised perform single cell-segmentation.
The brand new software is described in a brand new paper published within the journal iScience. The project was led by Assistant Professor of Biomolecular Engineering Ali Shariati and his graduate student Abolfazl Zargari. The model, called cGAN-Seg, is freely available on GitHub.
“The photographs that come out of our model are able to be used to coach segmentation models,” Shariati said. “In a way we’re doing microscopy with no microscope, in that we’re capable of generate images which are very near real images of cells by way of the morphological details of the one cell. The fantastic thing about it’s that once they come out of the model, they’re already annotated and labeled. The photographs show a ton of similarities to real images, which then allows us to generate recent scenarios which have not been seen by our model through the training.”
Images of individual cells seen through a microscope may also help scientists find out about cell behavior and dynamics over time, improve disease detection, and find recent medicines. Subcellular details resembling texture may also help researchers answer vital questions, like if a cell is cancerous or not.
Manually finding and labeling the boundaries of cells from their background is incredibly difficult, nevertheless, especially in tissue samples where there are a lot of cells in a picture. It could take researchers several days to manually perform cell segmentation on just 100 microscopy images.
Deep learning can speed up this process, but an initial data set of annotated images is required to coach the models — at the least hundreds of images are needed as a baseline to coach an accurate deep learning model. Even when the researchers can find and annotate 1,000 images, those images may not contain the variation of features that appear across different experimental conditions.
“You need to show your deep learning model works across different samples with different cell types and different image qualities,” Zargari said. “For instance should you train your model with top quality images, it is not going to have the ability to segment the low quality cell images. We are able to rarely find such a very good data set within the microscopy field.”
To deal with this issue, the researchers created an image-to-image generative AI model that takes a limited set of annotated, labeled cell images and generates more, introducing more intricate and varied subcellular features and structures to create a various set of “synthetic” images. Notably, they’ll generate annotated images with a high density of cells, that are especially difficult to annotate by hand and are especially relevant for studying tissues. This method works to process and generate images of various cell types in addition to different imaging modalities, resembling those taken using fluorescence or histological staining.
Zargari, who led the event of the generative model, employed a commonly used AI algorithm called a “cycle generative adversarial network” for creating realistic images. The generative model is enhanced with so-called “augmentation functions” and a “style injecting network,” which helps the generator to create a wide selection of top of the range synthetic images that show different possibilities for what the cells could seem like. To the researchers’ knowledge, that is the primary time style injecting techniques have been utilized in this context.
Then, this diverse set of synthetic images created by the generator are used to coach a model to accurately perform cell segmentation on recent, real images taken during experiments.
“Using a limited data set, we are able to train a very good generative model. Using that generative model, we’re capable of generate a more diverse and bigger set of annotated, synthetic images. Using the generated synthetic images we are able to train a very good segmentation model — that’s the major idea,” Zagari said.
The researchers compared the outcomes of their model using synthetic training data to more traditional methods of coaching AI to perform cell segmentation across various kinds of cells. They found that their model produces significantly improved segmentation in comparison with models trained with conventional, limited training data. This confirms to the researchers that providing a more diverse dataset during training of the segmentation model improves performance.
Through these enhanced segmentation capabilities, the researchers will have the ability to raised detect cells and study variability between individual cells, especially amongst stem cells. In the longer term, the researchers hope to make use of the technology they’ve developed to maneuver beyond still images to generate videos, which may also help them pinpoint which aspects influence the fate of a cell early in its life and predict their future.
“We’re generating synthetic images that may also be changed into a time lapse movie, where we are able to generate the unseen way forward for cells,” Shariati said. “With that, we wish to see if we’re capable of predict the longer term states of a cell, like if the cell goes to grow, migrate, differentiate or divide.”
