Polyomavirus Localized with Deep Learning Methods

Abstract

This case study shows how to apply a deep learning algorithm to detect specific particles in transmission electron microscopy images. The main difficulty consists in the confusing background where a mix of artifacts introduced by the negative staining method coexist with biological debris, and low contrast inherent to the acquisition process of the samples. Considering that a high-resolution 3D reconstruction of virus particles or other macromolecular structures needs thousands of particles, the recent advances in deep-learning methods makes the reach of a comfortable number of particles attainable with a high level of confidence. Deep learning algorithms based on convolutional neural networks can overcome the major issues after a train and validation stage, and subsequently process thousands of similar images in a short amount of time, with a few seconds required for each image, localizing all the particles of interest with very few mistakes, or missed particles. We applied one of such algorithms to a dataset of polyomavirus images containing about 390 particles and evaluate its performance against human curation, verifying that 88.9% of the dataset were processed with zero false positive detections.