A bespoke EMPS: Electron Microscopy Particle Segmentation dataset was constructed to serve as the training data for the authors’ work. It consists of 465 electron micrographs and their corresponding human-labeled ground-truth semantic instance segmentation maps, as well as the coordinates of the polygons drawn around each particle to construct these segmentation maps.

The figure shows 16 sample images and their segmentation maps and portrays qualitatively the diversity of particle sizes, shapes, textures, densities, and (grayscale) colors that exist in the dataset. Although not relevant for computing quantitative measures, the authors included several images where particles overlap each other with varying degrees of overlap, as this is common in the electron micrographs of nanoparticles. The third EM image in the second row of the figure above is an example of highly overlapping particles, while the particles in the fourth EM image of the third-row show minor overlap.

All images in the EMPS data set were mined from published scientific literature using the data-mining application programming interface (API) of Elsevier. Authors first used the Article Retrieval API to obtain the digital object identifiers (DOIs) of articles published between the years 2015 and 2020, which had the possibility of containing EM images. This was achieved using the search query “SEM−TEM−scanning electron microscopy−transmission electron microscopy.” Next, using the Object Retrieval API, authors iterated through the figures in these articles and obtained images at high resolution from any figure that contained one or more of the acronyms or phrases from their search query. This resulted in 34 091 images of figures, from which 788 were manually determined as suitable and set aside for postprocessing. It was often the case that EM images were part of a panel of several images in these figures. Thus, the EM images were cropped from these 788 figures, resulting in 962 potential images being labeled (many figures contained several relevant EM images). Authors annotated 465 of these images using the VGG Image Annotator (VIA). This consisted of drawing polygons around each individual particle in each image. Once the annotation process was completed, the authors finally assigned pixels to particle instances in each image by finding all pixels that were encapsulated by the polygon of each particle.

The figure presents some statistics of the images and particles in the EMPS dataset. It is evident that most images contain fewer particles, and only a few images contain many particles (on the left). Similarly, with particle sizes, most particles in the dataset are small, with the number of large particles dropping significantly as particle size increases (on the right).