AI at the edge – counting cells
SCS developed a deep learning algorithm that determines the number of cells on an image in the microscopy system. The AI was adapted for use "at the edge" - the neural network was optimised so that it can also run on a low-performance PC.
-
Problem definition
The development and growth of cells is monitored for many applications in biology, biotechnology and pharmaceutical technology. This can take place in different ways, for example due to different substances being tested. Until now, the cells on the corresponding microscopy images have been counted manually in a time-consuming process
-
SCS solution
SCS created an algorithm based on deep learning that automatically determines the number of cells in an image in the microscopy system. The algorithm was specially adapted for use on low-performance computers.
-
Added value
The customer can offer scientists a further feature that reduces the time they spend on research and development by many hours while maintaining the same quality and does not require any new purchases.
Project insights
The innovative zenCell Owl offers the customer a lightweight and stackable microscopy system for the incubator. It can be used to monitor the growth of cells in 24 different samples simultaneously. The monitoring and evaluation of the data is carried out conveniently on the PC. Both the degree of coverage and the number of cells can be determined automatically by the software. However, the development of the cell counting algorithm has special framework conditions.


On the one hand, the field of machine learning promises great success in the area of automatic image processing. Progress has also already been made in recognising, segmenting and counting cells. These should be utilised for the system.
However, the typical PC in a lab is not particularly powerful. Nevertheless, fast and reliable results are expected from the algorithm.


To solve this problem, SCS first trained an algorithm with fast hardware that reliably segments the cells. This then served as a reference for a resource-saving variant that still delivers comparably good results. Various cuts were gradually made and their effects on the accuracy and speed of the analysis were assessed.
This solution proved to be a reliable approach for such problems.

