Advancing Image Analysis through Machine Learning and Its Real-World Applications

Large volumes of microscopy images are generated every day, and effective analysis is essential for extracting meaningful information and understanding the properties of underlying materials. Manual analysis is time-consuming, and vulnerable to human error. It also lacks reproducibility, as analysts may make different judgments during key steps such as boundary identification for segmentation or particle size measurement. Machine learning has become a powerful tool in data science and computer vision, and its value in microscopy image analysis continues to grow. Machine learning-based advanced image analysis is gaining significant attention and adoption across the microscopy community. [1-4] In this presentation, we highlight the use of machine learning-based image analysis in an industrial laboratory environment. Particle size measurement is critically important because size directly influences key material properties such as surface area, processibility, and the viscosity of nanoparticle sols. Transmission electron microscopy (TEM) is widely regarded as the gold standard for obtaining accurate size information, supporting reliable product development and consistent manufacturing performance. However, TEM is limited in sampling throughput. High-volume measurement of thousands (even more) particles is increasingly required. We have developed a workflow to enable statistically robust counting and measurement of particle size and shape. Figure 1 exhibits a typical segmentation of 5 nm silica nanoparticles and the resulting statistical analysis, showing a relatively broad size distribution with an average particle size larger than the nominal 5 nm. This capability not only supports comparison or validation with techniques that generate bulk-sample averages, such as Small-Angle X-ray Scattering (SAXS), but also delivers essential particle-level insights needed for reliable material assessment and process control. Traditionally, heavy-metal stains such as RuO4 and OsO4 are applied to enhance contrast by targeting specific functional groups, enabling visualization of phase separation, particularly in low-contrast polymers. However, these staining reagents are toxic, vapor-based, and present environmental and user-safety concerns. In this presentation, we also demonstrate that stain-like quality TEM images can be generated for low-contrast polymer and adhesive samples by training a Cycle Generative Adversarial Network (CycleGAN) model on unstained and stained unpaired images. This approach can substantially enhance the contrast of phase separation in polymers without the need for staining, as shown in Figure 2. It eliminates repetitive staining steps and significantly increases throughput. (a) Representative image showing the segmentation of 5 nm silica nanoparticles, (b) Histogram of equivalent diameter illustrating particle counts and size measurements from the 5 nm silica nanoparticle sample. (a) An unstained TEM image used as input, (b) Corresponding predicted stained-like TEM image.