Researchers from SKOLKOVO Institute of science and technology has introduced a new algorithm that processes a biological image. It allows to precisely select individual biological objects from a complex photo. The results of the study are available on the website of preprints arXiv.org and will be presented as oral presentation at the conference on computer vision СVPR 2020.
For many biological studies require the analysis of the obtained images, for example, using microscopy. To automate this process quite difficult. For this we have to handle multiple layers and different objects at the same time, especially if we are talking about microscopic images in which the objects overlap each other, and the quality and clarity of the picture can be very low. To speed up such analysis by using the computer, which is capable through machine learning to quickly process large numbers of images and to isolate them from separate objects.
Researchers from Skoltech in the new work presented a new method for isolating biological objects such as cells, from complex images. It is based, scientists took the idea of simplifying complex problems of division of objects to more simple regression. This the authors were able to achieve with the introduction of additional “harmonic” signals in the neural network and automatic adjustment of parameters of these signals, the characteristic size and location of shared objects.
For the analysis, the researchers used four different types of images: images of plants, images with a large number of worms C. Elegans, microscopic images of bacteria E. Coli and the cultures of cancer cells HeLa. New machine learning technique consisted of two stages, the use of which allowed to improve the handling of images. In the course of work, scientists have trained a neural network in a separate data array for each type of images. This allowed us to analyze images of biological objects much more precisely the previously presented methods. Thanks to the algorithm, the researchers were able to identify the leaves of plants, worms, cancer cells and certain bacteria. The new method can find application in research and medical practice.
“the Main advantage of our approach is the ability to learn even in small samples. We hope that this machine learning technique will find applications not only in biological research but also in other industries where it is difficult to obtain large samples of marked training images,” says one of the researchers, Professor of Skoltech Victor Lempitsky.
