Correlating microscopy and ToF-SIMS images to cellulose size using deep learning

Abstract

In pharmaceutical formulations, chemically modified celluloses are used in several applications and it is often critical to have good quality control for these materials. In this thesis, spot testing after chromatographic separation has been evaluated as a method for material structure analysis. A highly non-linear correlation between material quality and spot appearance was expected and therefore supervised deep learning was used to model this relationship. Optical microscopy images were subjected to a pretrained resnet-18 image model to identify differences in chemical properties between spots. After suitable preprocessing, models could successfully be built to tell the difference between spots from early and late mass fractions. The cellulose fractions were also analyzed as parts of spots by ToF-SIMS. A 3D CNN model was trained from scratch. The model could successfully distinguish between fractions in this case as well.