A new automatic analysis tool for the determination of primary particle size from electron microscopy images: Application of the Cellpose software

To assess if an aggregated or an agglomerated material has to be considered as a nano-material, the size distribution of its constituent primary particles needs to be measured and the median diameter determined. To this end, the reference method uses either transmission or scanning electron microscopy to obtain images of the sample. The size of a significant number, usually a few hundreds, of primary particles are then measured manually. This task is highly time-consuming and subjected to operator bias. Some attempts have been made to automatize the size measurements. The algorithms and software available are generally successful at segmenting images of spherical objects with no or partial overlap but fail to properly segment irregular objects with strong overlap.

The advances made with deep learning algorithms are promising to solve the segmentation issues encountered so far on complicated samples. In this paper, we tested the open source deep learning Cellpose software on transmission and scanning electron microscope images of different samples to retrieve the median diameter of the primary particles and compare the results with both the manual and theoretical values. This software was chosen for its ease of use, its free availability and the fact that it is pre-trained, allowing the use of a limited set of training images.

For the samples used in this study, the quality of the segmentation was highly dependent on the number of objects on which the software model was trained, but a number of 500 to 1000 objects was enough to obtain good performances. The diameters measured using Cellpose segmentation are most of the time in agreement within 10% with the manual values. Interestingly, for scanning electron microscopy data, the results obtained with Cellpose are closer to the theoretical values when compared to the measurements obtained by hand, implying a smaller operator bias. If an uncertainty assessment still needs to be investigated for the diameters determined using Cellpose, this first attempt to use this software to segment electron microscope images of diverse samples is very promising and opens the possibility to fully automatize the identification of nano-structured materials.