Semi-automated image analysis of cellulose nanofibrils using machine learning segmentation and morphological thinning

Abstract

Quantitative width distributions of cellulose nanofibrils (CNFs) are difficult to obtain from microscopy when fibrils are highly branched and entangled, and when manual measurements rely on sparse sampling and subjective fibril selection. Here we present FACT (Fibril Analysis for Cellulose Technology), a semi-automated image-analysis framework that extracts length-weighted fibril width distributions from negative-contrast scanning electron microscopy (NegC-SEM) images. FACT uses machine-learning segmentation (either the ImageJ Weka plugin or a U-Net convolutional neural network) to generate binary CNF masks, then applies morphological thinning to obtain a one-pixel-wide skeleton. Local fibril width is calculated from the distance between each skeleton pixel and the segmented fibril boundary. Using idealized simulated geometries, hierarchical branched structures, and fixed-diameter wire micrographs, we identify practical operating limits for robust width statistics: high-contrast images, fibril aspect ratio greater than 10, and at least ~ 5 pixels across the fibril width. FACT was then applied to NegC-SEM images representing low- and high-branching CNF morphologies and compared with manual ImageJ measurements. Central tendencies were similar, while distribution shapes differed because FACT measures widths at many points along each fibril (effectively length-weighted), whereas manual analysis typically records one width per fibril (number-weighted); these outputs are therefore complementary rather than directly interchangeable. Once a trained image segmentation model is available, FACT analyzes each image in under 5 min, enabling higher-throughput morphology reporting.

Graphical Abstract

The alternative text for this image may have been generated using AI.