Efficient AI-supported quantification of network structures in SEM imaged human mucus: a detailed protocol

We present a detailed protocol for the training and application of an artificial intelligence (AI)-assisted workflow for automated analysis of pore and filament structures in scanning electron microscopy (SEM) images of human mucus from the respiratory tract and ileum. Our method leverages machine learning within the Fiji platform to classify image regions into pores and filaments, enabling precise and efficient quantification. The workflow includes preprocessing, pixel classification using a trained multilayer perceptron classifier, feature extraction, and data pooling. This approach significantly reduces analysis time compared with manual measurements while maintaining high precision. To validate our method, we compared manual and automated measurements on SEM images of airway and ileal mucus samples. High agreement was observed between datasets, with effect sizes indicating that AI-generated data variability is even less than human-to-human interobserver variation. In addition, we showed that increasing the counting frame area in automated analysis did not alter resulting distributions notably, confirming the robustness of the method. This AI-assisted tool supports high-throughput, comparative studies of biological network structures, enhancing accuracy and efficiency. Its application has the potential to accelerate research into structural and functional alterations of mucus in diseases such as cystic fibrosis and Crohn's disease, contributing to a deeper understanding of pathophysiology and the development of targeted therapeutic strategies.NEW & NOTEWORTHY Our study introduces a new artificial intelligence (AI)-assisted workflow for quantitative analysis of mucus network structures in scanning electron microscopy (SEM) images. By integrating machine learning with systematic uniform random sampling, this approach reduces variability and increases throughput compared with traditional manual methods. In our setup, the automated analysis was 40-120 times faster compared with manual analysis. The workflow is intended to be adapted by other biomedical research groups, particularly valuable for large-scale respiratory research.