Morphological classification of fine particles in transmission electron microscopy images by using pre-trained convolution neural networks

Morphological information on fine particles is essential for understanding their transport behavior in the ambient atmosphere and in the human respiratory system. More than 3000 transmission electron microscopy (TEM) images of fine particles were collected from ambient atmosphere and directly from various sources, such as diesel and gasoline engine exhaust, biomass burning, coal combustion, and road dust, and were then morphologically categorized into four major classes (spherical, agglomerate, polygonal, and dendrite). Pre-trained convolutional neural network (CNN) models (DenseNet169, InceptionV3, MobileNetV3Small, ResNet50, and VGG16) and traditional machine learning models (decision trees, random forests, and support vector machines) were trained using the classified particles. The fine-tuned CNN model (DenseNet169) having the deepest feature learning exhibited the best performance among the tested models, with an overall classification accuracy of 89% and an average per-class accuracy ranging from 84% to 97%. The reliable classification of thousands of images was performed within several minutes. The agglomerated class was the least misclassified because of its significantly different features from those of the other classes. The critical regions of the particles for classification decisions varied among the pre-trained models. Our results suggest that the pre-trained CNN models would be useful for the rapid morphological classification of a large number of fine particles.