However, the powerful data processing capabilities of modern deep learning models now provide techniques for classifying large amounts of image data. Assistant professor of biomolecular engineering Ali Shariati and PhD student Abolfazl Zalageri, along with several student researchers in Shariati’s lab, have developed and released a new deep learning model called ‘DeepSea’. It segments cells, tracks them, detects cell divisions and tracks cell lineage. DeepSea is described in detail in the following new papers: cell report methodis one of the most accurate tools of its kind.
DeepSea’s model training datasets, user-friendly software, and open-source code are available on the DeepSea website. Shariati and his team of researchers are already using it to make new discoveries about stem cell growth and division.
“The model is more efficient, has fewer parameters, and integrates both segmentation and tracking into easy-to-use software,” said Shariati. “With this software, we can train models for any cell type of interest, paving the way for future discoveries.”
Using time-lapse microscopy, which captures a series of images from a microscope over time, researchers can monitor single cells during an experiment to see if they differentiate (when a stem cell becomes a particular type of cell) or change shape. phenomena such as And it changes size over time. This will enable scientists to make new biological discoveries by measuring the dynamics of cellular biological phenomena at the single-cell level.
Once the scientist has collected the images, they must perform two main tasks: One is segmentation, distinguishing the boundaries of individual cells from each other and the background. And track or track the cells from one frame to the next. From that point on, researchers can further explore features such as size, shape, texture, and changes in movement and shape.
Manually classifying microscopic images is a tedious, time-consuming and ultimately computer-friendly task. This is where DeepSea comes into play. This efficient deep learning model allows him to perform segmentation in less than a second and track cells with 98% accuracy.
Enabling the software to detect cell division was a particularly unique and challenging aspect of this project. There are very few other situations in which artificial intelligence and computer vision would have to track the transformation of one object into another.
“This is a very unusual problem in object tracking,” Shariati said. “If you want to track something like a car, it’s moving around, and you can use machine learning and computer vision to track it’s movement. It was a fundamentally new problem that we needed to solve, and we were able to solve it.”
DeepSea is a generalizable model and can be used to track many types of cells. It uses a modified version of the popular model 2D-UNET with significantly reduced parameters to achieve both high speed and accuracy.
“We have compared our model to several of the best cell segmentation models. Currently, our model shows the best results in terms of accuracy and speed, especially for these cell types,” he says. Zalagueri, who holds a PhD in electrical and computer engineering, said: A student in Shariati’s lab who led the creation of the software.
The researchers trained DeepSea using a dataset of images of cells that were manually segmented from the background, but the images were often poorly contrasted and cell bodies were difficult to discern, making them time-consuming. It has become such a process. To assist in this process, the team developed another software tool to aid in cropping, labeling, and editing microscopic images of cells. This tool is also available on his DeepSeas.org.
The training dataset contains lung, muscle, and stem cell images, which means DeepSea achieves high accuracy across different cell types. Future versions of the model may add more cell types.
Researchers used DeepSea to study the size control of embryonic stem cells, the basis of multicellular life and capable of differentiating into all other cell types. Their new discovery is that embryonic stem cells, known to divide abnormally quickly, regulate their size so that smaller cells spend a long time growing before giving rise to the next generation of cells. brought
“We found that if embryonic stem cells were born small, they knew to some extent that they were small, so they spent a lot of time growing before dividing again,” said Shariati. “I don’t know why or exactly how this happens, but at least the phenomenon exists.”
In the future, the researchers will apply existing software to collect data to study the spatial relationships between cells and how cellular features are organized in 3D patterns to form structures. is.
The researchers also aim to solve bottlenecks they have noticed when using deep learning models, such as the lack of labeled images of cells used to train the models. They use a class of machine learning frameworks called generative adversarial networks (GANs) to generate new synthetic data, images of cells that have already been annotated to reduce the time it takes to create labels. I am planning to create Researchers would then be able to create large libraries of datasets for any cell type of interest with minimal human involvement.
