Automatic SARS-CoV-2 Segmentation in Electron Microscopy Based on Few-shot Learning

Abstract

Due to the advantages of direct visualization and high resolution, transmission electron microscopy (TEM) technology has been widely used in the morphological identification of viruses. With the development of artificial intelligence (AI), there have been some studies on automated TEM virus identification using deep learning. However, to achieve effective virus identification results, a large number of high-quality labeled images are required for network training. In this work, we propose an automatic virus segmentation method based on few-shot learning. We use the Chikungunya virus, Parapoxvirus and Marburg virus, etc. to construct a pre-training virus dataset and train an attention U-Net-like network with an encoder module, relationship module, attention module and decoding module to realize severe acute respiratory syndrome coronavirus type 2 (SARS-CoV-2) segmentation using few-shot learning. The experiment shows that the proposed few-shot learning methods yield 0.900 Dice and 0.828 Jaccard in 1-shot, 0.903 Dice and 0.832 Jaccard in 5-shot, which demonstrates the effectiveness of our method and outperforms other promising methods. Our fully automated method contributes to the development of medical virology by providing virologists with a low-cost and accurate approach to identify SARS-CoV-2 in TEM.