Automated SEM Acquisitions and Segmentation With AI

Electron microscope parameters are routinely tuned prior to acquiring data to alter image contrast, noise, lateral and depth resolution and more. Depending on the desired results, microscopists select the correct combination of parameters that will produce images that they consider satisfactory. Ultimately, tuning parameters can be daunting when microscopists are without a certain expertise and results are not reproducible since selected parameters may vary from one microscopist to another. Furthermore, there is no guarantee that acquired images will be successfully segmented by algorithms during post processing steps. Segmentation is often necessary for quantitative analysis of any SEM image and manual labelling is extremely time consuming and error prone. Image processing software offer multiple segmentation algorithms to avoid manual labelling, and their efficiency is closely related to the contrast, noise, and resolution of acquired images. Suitable microscope parameter selection will generally determine the outcome of segmentation success. We propose integrating artificial intelligence (AI) within SEM acquisition workflows to optimize images for efficient segmentation. Automated microscope parameter selection with trained regression models will considerably improve data acquisition with the SEM for quantitative analysis. First, it will ensure the success of segmentation algorithms by producing images for that purpose, second, it will allow reproducibility, providing an unbiased selection of microscope parameters with observed consistency throughout images, and third, it will make SEM workflows more accessible by eliminating the need for an expertise on electron material interactions to produce desirable images. Electron microscope parameter prediction models are trained with data generated using Monte Carlo simulations. A python-based script with the McXRay [1] plugin in Dragonfly creates multiple back scattered electron (BSE) images from initially labelled samples, by varying both the beam energy and the probe current as shown in Figure 1. The samples used to generate the training data have a predetermined composition, for experiments, platinum nanoparticles on carbon nanotubes are used. Over 5000 simulations are generated, each attributed a DICE score [2] computed by comparing already labeled virtual samples, with simulated BSE images segmented with simple Otsu thresholding. Using simulations on virtual samples at different magnification, with nanoparticles of various shapes and sizes and positioned at different depths will intentionally diversify training data, providing the regression model with as much variety as possible and improving its prediction accuracy [3]. Models are also trained with a user specified segmentation algorithm, implemented in the described python script. The model, with architecture illustrated in Figure 2, is originally trained to predict only two SEM parameters, the beam energy and probe current, as a proof of concept, and due to their considerable influence on image characteristics, as demonstrated in Figure 1. The beam energy will impact contrast, lateral and depth resolution and the probe current will determine the signal to noise ratio and resolution. Once properly trained, the model will output the appropriate parameters (beam energy and probe current) to be set on the microscope. Conclusively, integrating AI into SEM workflows reliably acquires data for segmentation, with minimal effort from microscopists. Automated McXRay simulations of platinum nanoparticles (white or pink) on carbon background (black or green) generated by varying both the beam energy and probe current, with their respective segmentations using Otsu thresholding. a. 3 keV and 0.2 nA with dice score: 0.70 b. 3 keV and 100 nA with dice score: 0.87 c. 20 keV and 100nA with dice score: 0.83 d. 20 keV and 0.2 nA with dice score: 0.25 Regression model architecture for predicting beam energy and probe current for accurate segmentation. The inputs include a SEM image (or simulation from McXRay) and a DICE score to account for segmentation accuracy. Users can train the model with any data of the same composition and with a selected segmentation algorithm.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.