Cell counting is extremely important in research, medicine, and even environmental monitoring where scientists use it to track cell growth, a person’s health, or monitor plankton levels in oceans or bacteria in a water sources.
But scientists who have used a hemacytometer, a specialized laboratory device used for manual cell counting, might tell you how challenging it can be to accurately determine cell numbers. This is because the hemacytometer consists of a thick glass slide with a rectangular indentation that creates a counting chamber. The chamber is divided into grids or squares with known dimensions, allowing for accurate cell counting and concentration calculations. It can be quite a challenge to figure out the number of cells in those tiny spaces.
“Manual cell counting is a tedious task,” explained Yudong Zhang, professor of School of Computing and Mathematical Sciences, University of Leicester in an email. “It requires operators to count cells in the small grids of a counting board under a microscope. The grids on the counting board are divided into tiny sections, making it easy to make counting errors. Moreover, performing such a concentration-demanding task for a prolonged period can also have an impact on the operator’s physical well-being.”
Zhang, therefore, wondered whether in this age of AI and automation if something more could be done to alleviate the burden of manual counting methods, which are often time consuming, labor intensive, and susceptible to human error.
“Last year, while tutoring my cousin for his high school assignment, I came across a question about using a blood cell counting board to count cells,” said the study’s co-author, Lijia Deng. “It made me curious if there were AI technologies available for this purpose. After conducting a bit of research, I found that there were opportunities to improve existing cell counting methods.”
Alongside colleagues, Shuihua Wang and Qinghua Zhou from the same university, the team set out to alleviate the burden of manual counting. To do this, they created an innovative automated detection method powered by AI.
An AI-based approach
Automated cell counting methods are not completely absent from these fields. “However, mainstream instruments are based on the Coulter Principle, which is the detection and measurement of changes in electrical resistance produced by a particle or cell suspended in a conductive liquid,” explained Zhang. “These instruments do not provide visual feedback, and cell morphology often reflects important information, such as the differences between cancer cells and normal cells.”
In a recent study published in Advanced Intelligent Systems, the team unveiled a revolutionary deep learning network they called Spatial-based Super-Resolution Reconstruction Network (SSRNet), which was spearheaded by Deng. “This network predicts cell counts and segments cell distribution contours with remarkable precision,” said Zhang.
Using this method, the cell sample is captured as an image which is then processed to enhance the clarity of the cells against the image’s background. The image is then fed to the AI counting system, which generates the cell count and distribution within the image.
“This AI-based approach can quickly predict the number and distribution of cells with just a single image,” said Zhang. “The principle of this method lies in the convolutional neural network’s focus on cell features, enabling the prediction of cell count and distribution.”
Traditionally, AI uses artificial neural networks — computational models inspired by the structure and function of the human brain — to perform tasks and learn from encountered situations. “Training any neural network model requires rich datasets,” added Zhang. “And there is a lack of sufficient, annotated datasets in the field of cell imaging.”
The team therefore took a different approach to overcome the lack of data needed to train their model, instead using it to predict the overall quantity and distribution regions to accomplish the task of cell counting.
They did this by taking advantage of a concept called upsampling, which is a technique used to increase the resolution or sampling rate of digital data. It involves taking existing digital samples and adding extra samples in between them to create a higher-resolution version of the original data.
” The traditional method is to use purely mathematical methods, which introduce new pixel values due to mathematical calculations,” explained Deng. “Although these new pixels make the image appear clearer, they can affect the prediction of quantity. Our method uses artificial intelligence to predict new pixels, reducing the potential system errors caused by mechanical calculations, improving counting accuracy, and also achieving the performance of traditional methods in clarity.”
“It’s like rolling out the dough after fermentation — our approach doesn’t introduce new pixels out of thin air; each new pixel is inferred from existing ones,” Deng continued. “Compared to purely mathematical methods, our approach ensures better consistency between the upscaled image and the original image in terms of features. Additionally, the larger the scaling factor, the more apparent the advantages become.”
Beyond just cell counting
There was also the added challenge of ensuring their AI system could be used anywhere, even in regions with limited computing resources. “To help popularize our AI model and make it available to labs that may lack advanced computing resources, we made our neural network model extremely lightweight so that its running memory read and write consumption is only 1/10 of a traditional AI model.”
The innovative features of their AI model will allow it to find application beyond just medicine and biology, promising to unlock new possibilities in various industries. As proof-of-concept, the team demonstrated how this model could be used to count the number of sesame seeds on a piece of bread.
Sesame counting was done just for fun, say the team, it has no practical significance but demonstrates the method’s sophistication and speed, which could one day be applied to more advanced applications, including cell counting, among others. “For example, we could eventually use aerial photography to remotely capture the breeding population of penguins to understand their population size, which avoids human interference with animals,” explained Deng.
“This method represents a significant leap forward in the field of cell counting,” said Zhang. “By leveraging the power of AI and innovative spatial-based super-resolution reconstruction techniques, this approach offers unprecedented precision and efficiency in predicting cell numbers and distributions, which can help fight against infectious diseases.”
With its potential, this advancement promises to streamline processes, reduce human error,. As the research continues, further refinements and applications of this AI-powered method are expected to reshape the landscape of cell analysis, ultimately benefiting countless individuals and facilitating scientific progress.
Reference: Lijia Deng, Qinghua Zhou, Shuihua Wang, Yudong Zhang, Spatial-Based Super-resolution Reconstruction: A Deep Learning Network via Spatial-Based Super-resolution Reconstruction for Cell Counting and Segmentation, Advanced Intelligent Systems (2023). DOI: 10.1002/aisy.202300185
Feature image credit: Scott Webb on Unsplash