Application of Artificial Neural Networks for Analysis of Ice Recrystallization Process for Cryopreservation

Cryomicroscopy is one of the main techniques to visualize freezing and thawing events taking place during cryopreservation of cells, native and artificial tissues with the ultimate goal to provide cell- and tissue-specific cryogenic preservation. However, there is currently no unified software tool for the automated analysis of ice recrystallization kinetics for a variety of cryoprotective agents used in the cryobiological practice. In this regard, we focused on the particular aspect of image analysis in the course of ice recrystallization, i.e. the possibility of using a neural network for the segmentation of ice crystals during isothermal annealing. In the work, the U-Net deep neural network was used for segmentation of ice crystals on cryomicroscopic images. Using 100 images as training set, the resulting accuracy of ice crystal segmentation was about 74% on the test sample (30 images). The obtained results show the possibility of segmentation of ice crystals in cryomicroscopic images taking into account the overlapping of intensity levels of an object and background. Further improvement of the model through the use of an additional training data as well as augmentation techniques is required to more efficiently validate this approach.