Cryogenic electron microscopy (Cryo-EM) produces high-resolution 3D images at angstrom levels used by researchers across a broad range of fields including structural biology, life Science, materials science, nanotechnology, semiconductors, energy, environmental science, and food science. Advancements in microscopy hardware enable production of 2D and 3D micrographs with near Angstrom resolution but require exponentially increasing data processing and storage capability. Images generated by cryo-EM are visually noisy, and each project can produce more than 100,000 images and take weeks to arrive at one viewable 3D structure. Many steps in the cryo-EM workflow require manual intervention and analysis that can take several weeks and result in errors due to user bias, time waiting and user fatigue. Current image processing and data analysis solutions are not wellintegrated, requiring extensive manual user involvement and long wait times before assessing image quality. Here we describe our development of machine learning models for automation of single particle classification during cryo-EM image processing with repeatable accuracy levels and integrated into the cryo-EM workflow for easy deployment with a new machine learning platform, called CryoDiscovery (Figure 1). We tested several Convolution Neural Network (CNN) designs for ML training and inference using a private set of over 20,000 images and metadata files. CNN architectural considerations include network depth, activation function and hyperparameters. Our CNN processed image data via a layered approach, iteratively through repeated transformations (in the "hidden" layers) to extract features before classifying them (in the "output" layer) 2-D and 3-D class selection. CNN models were trained using image data found in mrcs files, non-image metadata found in star files, and image annotations (ground truth) found in selection files using a computer with a dual socket 2nd gen Intel Xeon® CPU (8 cores each) with 4 NVIDIA 2070-Ti GPUs and 96GB of physical memory. Data preparation was conducted by trained researchers prior to ML training, and consisted of image retrieval, resolution normalization, image augmentation, and metadata selection. Verification of our models was done by analyzing maximum prediction accuracy with low variance, and false negatives to minimize misclassification of good data, and the impact of using metadata to improve model prediction accuracy. Model boosting was used to generate strong prediction algorithms and more consistent results from multiple simple models [1]. Three models were trained sequentially and used for inferencing, as shown in Figure 2. The third model was used when the first two models disagreed for the production data. Fourier Shell Correlation vs. Resolution (1/Aº) [2] was used to verify that the resolution (at threshold) meets published results. Secondly, we calculated the Mean Square Error (MSE) of 2D predicted images vs. ground truth images to provide a leading indicator of 3D model differences. Lastly, we examine Struct...
Machine learning has become a great attention to find optimize solutions in different areas and is anticipated to play a vital role in our upcoming technologies. This paper presents a comprehensive review on basic optimization algorithms for micro-strip patch antenna design using machine learning. Classification of machine learning based algorithms: deterministic, stochastic and surrogate model assistant is discussed. Further machine learning models training for optimizing output and for prediction of antenna parameters is presented in this paper. This paper is useful to the readers who work on a particular antenna using the Machine Learning Techniques.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.
hi@scite.ai
10624 S. Eastern Ave., Ste. A-614
Henderson, NV 89052, USA
Copyright © 2024 scite LLC. All rights reserved.
Made with 💙 for researchers
Part of the Research Solutions Family.