Deep learning methods are popular for cryo-electron microscopy (cryo-EM) particle picking, but their generalization to unseen datasets is unpredictable, and they often forget previously learned knowledge. This paper introduces EPicker, an exemplar-based continual learning approach that accumulates knowledge from new datasets without catastrophic forgetting. EPicker uses a dual-path network and a novel loss function to achieve this, enabling it to continuously improve particle picking performance and expand its ability to identify various biological objects (proteins, vesicles, fibers).
Publisher
Nature Communications
Published On
May 05, 2022
Authors
Xinyu Zhang, Tianfang Zhao, Jiansheng Chen, Yuan Shen, Xueming Li
Tags
deep learning
cryo-electron microscopy
particle picking
continual learning
biological objects
catastrophic forgetting
exemplar-based
Related Publications
Explore these studies to deepen your understanding of the subject.