Semi-automated image analysis of Cellulose Nanofibrils using Machine learning segmentation and Morphological thinning

Abstract

Reliable and rapid morphology measurement of cellulose nanofibrils (CNFs) with a high level of branching and entanglement is crucial for quality control, grade definition, and investigating morphology-performance relationships in various applications. An image analysis framework, Fibril Analysis for Cellulose Technology (FACT), which utilizes machine learning (ML) segmentation and morphological thinning, was developed to measure the fibril width distribution of cellulose nanofibers (CNFs) from negative contrast scanning electron microscopy (NegC-SEM) images. The high-contrast and wide magnification range of NegC-SEM imaging enabled the capture of micro- and nanoscopic hierarchical branching structures of CNFs. Two ML approaches [Weka and U-Net] were used to create detailed binary segmentation of grayscale NegC-SEM images, critical for the width analysis. Morphological thinning was applied to the binary image to produce a 1-pixel-wide skeleton of the CNF fibril structure. Subsequently, the distance between the skeleton and the original fibril edge was used to calculate fibril width. The FACT framework was optimized and validated with idealized geometric and hierarchical branched structures. FACT effectively performed segmentation, skeletonization, and fibril width measurement of these CNF morphologies. FACT width results were comparable with manual measurements. In the manual method, a single measurement is made per fibril. In contrast, FACT simultaneously makes multiple measurements along each fibril within the entire CNF branched network structure. The advantage of FACT is that complicated branching and network CNF structures can be measured without imparting any analyst bias in fibril selection and measurement. Additionally, once the ML model is trained, each image can be analyzed in under 5 minutes.