A Novel Clustering Method Using Variational Autoencoder with Reliable Sample Decision and Balanced K-Means++ for Single-particle Cryo-EM Images

Single-particle cryo-electron microscopy (cryo-EM) is one of the most popular technology in the field of biology molecular structure determination. Clustering for Cryo-EM particle images is very important in the structure reconstruction process, which significantly affected the reconstruction resolution. Because the signal-to-noise-ratio (SNR) of cryo-electron is extremely low, it's a challenge to improve the clustering accuracy. In this paper, we proposed a novel clustering method that combined a variational autoencoder with reliable sample decision (ReVAE) and balanced K-means++ (BK-means++). ReVAE projects cryo-EM images into low-dimensional latent variables, and BK-means++ is applied to cluster latent variables. Training of ReVAE and clustering of latent variables by BK-means++ are performed jointly and iteratively. The experimental results showed that ReVAE with BK-means++ achieved state-of-the-art results compared to traditional cryo-EM particle images clustering methods.