(510a) Detection and Classification of Chiral Inorganic Particles in Electron Microscopy Images Using Generalizable Deep Learning Algorithms

A chiral object has two mirror-image forms that are non-superimposable in three dimensions. Chirality plays a crucial role in chemistry, biology, and pharmacology, as most of the important biomolecules are chiral (amino acids, proteins, DNA). In 1998, it was discovered that chiral nanostructures could be chiral. Chiral inorganic nanostructures have distinct fundamental importance and are essential for further developing chemical, pharmaceutical, environmental, and biomedical technologies. To date, many researchers are focused on the establishment of the correlations between chiroptical and morphological properties of these materials by using circular dichroism and electron microscopies. A thorough investigation of electron microscopy images requires accumulating many images with their in-depth analysis, which is tedious and the 'manual' image processing is subject to experimentalist bias. Currently, there is a need for novel methods for the structural characterization of chiral nano- and micron-scale inorganic structures.

Here we developed an approach for synthesizing large sets of realistic scanning electron microscopy (SEM) images of chiral microscale helices of bowtie shape based on very small original SEM image sets (~200 SEM images). We used SEM images of a diverse pool of bowties, which allowed the training of state-of-the-art neural network models with the ability to detect morphological properties of chiral structures of various sizes. We tested the method by generating 10,000 images of bowtie-shaped particles with different chirality and training the YOLOv5 model to differentiate between right and left structures. As a result, this algorithm can reliably identify and classify chiral bowtie-shaped particles with accuracy as high as 94.4%. Furthermore, after training on bowtie particles, this model can successfully recognize other chiral shapes with different geometries without re-training. These findings indicate that deep learning techniques can potentially replicate the visual analysis of chiral objects, which opens up a path to other computational methods capable of accurate automated analysis of a wide range of chiral features at different scales and their implementation in materials discovery for photonics and medicine.